Universidad de Sevilla Estudio, Diseño, Implementación y Test de Retinas

Universidad de Sevilla Estudio, Diseño, Implementación y Test de Retinas
Universidad de Sevilla
Departamento de Electrónica y Electromagnetismo
Estudio, Diseño, Implementación y Test de Retinas
Visuales VLSI Sensoras de Contraste Espacial y
Temporal
Memoria presentada por:
Juan Antonio Leñero Bardallo
para optar al grado de Doctor por la Universidad de Sevilla
Sevilla, 2010
Estudio, Diseño, Implementación y Test de Retinas
Visuales VLSI Sensoras de Contraste Espacial y
Temporal
Memoria presentada por:
Juan Antonio Leñero Bardallo
para optar al grado de Doctor por la Universidad de Sevilla
Sevilla, 2010
Directores:
Dr. Bernabé Linares Barranco
Dra. Teresa Serrano Gotarredona
Tutora por la Universidad de Sevilla:
Dra. Adoración Rueda Rueda
Departamento de Electrónica y Electromagnestismo
Universidad de Sevilla
A mis padres
v
AGRADECIMIENTOS
Llegados a este punto, me gustaría agradecer a mis directores de tesis,
Bernabé Linares y Teresa Serrano, su ayuda, tiempo y dedicación durante el
tiempo de elaboración de la tesis. Sin sus valiosos consejos, la redacción de este
documento no hubiera sido posible.
También me gustaría agradecer a mis padres y hermanos su apoyo y comprensión durante estos años.
He de mencionar al programa JAE (Junta para la Ampliación de Estudios) del CSIC para la formación de investigadores, gracias al cual he obtenido
financiación para mi formación como doctor, así como la oportunidad de realizar
una estancia de dos meses en el Departamento de Informática de la Universidad
de Oslo, tutelado por el profesor Philipp Häfliger.
Son varias las personas que han colaborado en la realización de la tesis.
He de citar al grupo de Arquitectura de Computadores (ATC), de la Universidad
de Sevilla, por las PCBs y el software suministrados para el test de los sensores
diseñados; a Philipp Häfliger por la placa de test y los montajes para lentes suministrados; a Tobi Delbrück por el software jAER y su retina de contraste espacial sumnistrada; y a los revisores anónimos que han contribuido a mejorar la
calidad de los artículos que hemos publicado fruto del trabajo de esta tesis.
Por último, me gustaría citar a mis compañeros del CNM en Sevilla, con
los que he trabajado e intercambiado conocimientos durante la elaboración de mi
tesis. También, gracias a ellos, mi estancia en el CNM durante estos cuatro años
ha sido francamente grata.
vii
viii
Índice de Contenidos
Sistemas de Visión AER Sensibles al Contraste ................................................ 13
1 Introducción . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Sistemas de Visión Convencionales y Biológicos . . . . . . . . .
1.2 Extracción del Contraste Temporal. . . . . . . . . . . . . . . . . . . .
1.3 Extracción del Contraste Espacial . . . . . . . . . . . . . . . . . . . .
2 Corrientes de Polarización Programables (The I-Pot System) . . .
2.1 Introducción. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Arquitectura de los I-Pots . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Resultados Experimentales . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Conclusiones sobre los I-Pots . . . . . . . . . . . . . . . . . . . . . . . .
3 El Nuevo Método de Calibrado (Newcalib) . . . . . . . . . . . . . . . . .
3.1 Introducción. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Nuevo Sistema de Calibrado (Newcalib). . . . . . . . . . . . . . . .
3.3 Uso de Circuitos Translineales para el Calibrado . . . . . . . .
3.4 Estrategias para el Calibrado . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Resultados Experimentales . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6 Conclusiones sobre el Sistema de Calibrado . . . . . . . . . . . .
4 La Retina AER de Contraste Espacial . . . . . . . . . . . . . . . . . . . . .
4.1 Introducción. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Píxel de Contraste Espacial de Boahen. . . . . . . . . . . . . . . . .
4.3 Nuevo Diseño del Píxel . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4 Resultados Experimentales . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Conclusiones sobre la Retina AER de Contraste Espacial . .
5 La Retina AER de Contraste Temporal . . . . . . . . . . . . . . . . . . . .
5.1 Introducción. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Arquitectura del Píxel y Layout . . . . . . . . . . . . . . . . . . . . . . .
5.3 Circuito Externo de Control de la Ganancia. . . . . . . . . . . . .
5.4 Resultados Experimentales . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Conclusiones sobre la Retina AER de Contraste Temporal .
6 Conclusiones y Líneas Futuras de Investigación . . . . . . . . . . . . .
13
13
14
14
15
15
15
18
18
20
20
20
21
23
24
26
26
20
28
29
31
34
35
35
35
38
41
43
44
Bibliografía......................................................................................................... 47
ANEXOS ........................................................................................................... 55
ix
Contents
OUTLINE ....................................................................... 61
CHAPTER 1
Introduction...................................................................... 63
1.1 Conventional Vision Systems Based on Frames . . . . . . . . . . . .63
1.1.1 Charged-Coupled Devices . . . . . . . . . . . . . . . . . . . . . . . . . 65
1.1.2 The CMOS Sensor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
1.1.3 Color Cameras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
1.2 Bio-inspired Systems and Neuromorphic Engineering . . . . . . . 68
1.3 The Biological Retina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
1.4 AER Vision Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
1.4.1 Adaptive Temporal and Spatial Filtering . . . . . . . . . . . . . . 74
1.4.2 Spatial Contrast Extraction . . . . . . . . . . . . . . . . . . . . . . . . 74
1.4.3 Temporal Contrast Extraction . . . . . . . . . . . . . . . . . . . . . . 75
1.4.4 Time To First Spike Imagers. . . . . . . . . . . . . . . . . . . . . . . . 75
1.4.5 Luminance Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
1.5 New Developed AER Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . 76
CHAPTER 2
Biasing for Neuromorphic Arrays: The I-Pot System...... 79
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 I-Pots Circuit Architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Temperature Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
CHAPTER 3
The New Calibration System ............................................ 89
3.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 The New Calibration System: Newcalib . . . . . . . . . . . . . . . . .
3.2 MOS Transistor with Digitally Adjustable Length . . . . . . . .
3.3 Translinear Circuits for Tuning. . . . . . . . . . . . . . . . . . . . . . .
3.4 How to Optimize the Calibration Ranges . . . . . . . . . . . . . . .
3.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
CHAPTER 4
79
80
85
86
87
89
90
90
93
95
98
99
The AER Spatial Contrast Sensor .................................. 103
4.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.2 Prior AER Mismatch-Calibrated Unipolar Spatial Contrast AER
Retina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
x
4.3 Boahen Spatial Contrast Pixel . . . . . . . . . . . . . . . . . . . . . . . .
4.4 Improved Signal-Spatial-Contrast Pixel . . . . . . . . . . . . . . . . .
4.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.1 Pixel Frequency Range . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.2 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.3 Contrast Step Response . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.4 Contrast Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.5 Contrast Thresholding . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.6 Latency Characterization . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.7 Natural Scenes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.8 TFS Output Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.9 Power Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
CHAPTER 5
The AER Temporal Contrast Sensor .............................. 137
5.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 The Pixel Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Automatic Gain Control Block . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.1 Uniformity of Response. . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.4 Bandwidth Measurements. . . . . . . . . . . . . . . . . . . . . . . . .
5.4.5 Latency Characterization . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.6 Dynamic Range. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.7 Example Data and High Speed Applications . . . . . . . . . .
5.4.8 Power Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
CHAPTER 6
106
115
115
118
118
120
122
123
126
128
128
131
132
137
138
142
144
147
155
164
165
167
171
172
Conclusions .................................................................... 175
6.1 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
6.2 Achievements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
6.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
A
Abbreviations and Mathemathical Conventions ................................ 179
Bibligraphy..................................................................................................... 181
xi
xii
SISTEMAS DE VISIÓN AER SENSIBLES AL
CONTRASTE
1 Introducción
1.1 Sistemas de Visión Convencionales y Biológicos
En los últimos años hemos podido apreciar un crecimiento exponencial en las ventas de sistemas de visión. Dicha tendencia no parece detenerse y es, en parte, debida al auge de los sistemas
de telefonía móvil. Probablemente, cualquier persona está familiarizada con conceptos como el
número de Megapíxeles o la resolución de una cámara. Ello da una idea de la importancia que
tales dispositivos tienen hoy en día.
Los sistemas de visión convencionales están basados en frames. Un frame no es más que
una matriz bidimensional en la que cada elemento representa a un píxel y almacena una información relacionada con el nivel luminosidad en la región espacial que ocupa dicho píxel. Tales
matrices se generan de forma periódica y continua a lo largo del tiempo. Ello conlleva una serie
de limitaciones importantes. La primera de ellas es que hay tiempos muertos mientras se lee la
información asociada a cada píxel para generar el frame. Los cambios que ocurran en la escena
durante esos intervalos de tiempo no serán detectados. Además, los sistemas de visión convencionales no suelen realizar un procesamiento de la información. Simplemente transmiten valores de
intensidad luminosa, que puede ser muy redundantes en el caso de que no haya cambios en la
escena. Otras limitaciones que encontramos son el hecho de todos los píxeles tienen la misma
ganancia (lo cual conlleva problemas en el caso de haya regiones con niveles de luminosidad distintos) y bajo rango dinámico de operación.
Si comparamos las prestaciones de los sistemas de visión convencionales con los biológicos,
vemos que los segundos tienen claras ventajas sobre los primeros [4]. Aunque existen sistemas de
visión artificial que ofrecen muy buenas prestaciones, normalmente su funcionamiento se restringe a condiciones de iluminación muy específicas y controladas. La retina humana es capaz de
ofrecer un funcionamiento excelente en las más diversas condiciones de luminosidad. Además
está optimizada para extraer y enviar la información relevante de la escena al cerebro.
Por ello, parece razonable tratar de imitar el funcionamiento de los sistemas biológicos para
aprovechar sus ventajas inherentes. Éste es el objetivo de la disciplina conocida como ingeniería
neuromórfica. El auge de los sistemas VLSI (Very Large Scale Integration) posibilita incluir grandes poblaciones de píxeles en un chip. Los sistemas de visión biológicos son extraordinariamente
complejos y están estructurados en capas de neuronas. Las neuronas de cada capa están masiva-
13
Sistemas de Visión AER
mente interconectadas con las siguientes y se encarga de realizar un determinado procesamiento
de la información. Al contrario que los sistemas de visión convencionales, la información no se
transmite en forma de frames de forma periódica. La transmisión de información es continua a lo
largo del tiempo y ello posibilita que sea mucho más rápida. Además, la información transmitida
no es redundante. Sólo se transmiten los cambios espacio-temporales [4].
Si deseamos implementar tales sistemas en silicio, es necesario usar medios de comunicación eficientes entre las distintas capas de neuronas. Ello supone una limitación cuando se desean
interconectar grandes poblaciones de neuronas porque el número de conexiones que pueden realizarse en silicio es bastante limitado. En ese sentido, el método AER (Address Event Representation, [1]-[2]) es altamente eficiente para interconectar grandes poblaciones de neuronas.
Tal como mostraremos más adelante, las retinas AER tienen mayor ancho de banda, menor
latencia, mayor rango dinámico y menor consumo que los sistemas basados en frames. Ello justifica, la realización de esta tesis.
El objetivo de la tesis consistió en diseñar dos sensores AER de visión. El primero de ellos
es capaz de detectar el contraste espacial y el segundo, el temporal.
1.2 Extracción del Contraste Espacial
Desde que el primer sensor AER, fuera diseñado por Mahowald y sus colegas, [6], en 1992,
han aparecido varios sensores AER capaces de detectar el contraste temporal. Es de especial relevancia el trabajo [44] presentado por K. Boahen y A. Andreou en 1996. Aunque las salidas no
eran AER y estaban basadas en frames, los autores implementaron un sistema neuromórfico capaz
de detectar el contraste espacio-temporal. La retina era capaz de captar imágenes con aceptable
calidad, presentaba un consumo muy reducido y realizaba un AGC local, lo cual dotaba al sensor
de un gran rango dinámico. Sin embargo, tenía una serie de limitaciones, como el el alto mismatch. En particular, el mismatch ha sido y es uno de los campos de batalla de los sensores de contraste espacial. Posteriormente, los autores implementaron una retina similar a la enterior con
salida AER [21]-[22]. Sus principales limitaciones seguían siendo el elevado mismatch junto con
el consumo innecesario de ancho de banda cuando no se detectaba contraste espacial.
Una de las tareas a realizar durante este tiempo consistió en la implementación de un sensor
sensible al contraste espacial [23] que solventara las limitaciones antes comentadas. En concreto,
partiendo del trabajo de Boahen, se diseñó un nuevo sensor AER robusto frente al mismatch y que
sólo consumía ancho de banda cuando se detectaba contraste espacial.
Para hacer el sensor robusto frente al mismatch, se incorporó y diseñó un nuevo sistema de
calibrado. Además, el sensor incorpora un modo opcional de funcionamiento (time-to-first spike),
basado en las teorías de Thorpe [11]. Este método de funcionamiento combina las ventajas del
procesamiento basado en frames y el de los sistemas AER. En él, tras un reset global los píxeles
que detectan información más relevante disparan primero. Luego, tras un periodo de escaneo,
todos los píxeles son reseteados y el ciclo comienza de nuevo. El cerebro humano realiza un procesamiento de la información similar cuando por ejemplo, giramos la cabeza de forma repentina y
miramos hacia un lugar distinto.
1.3 Extracción del Contraste Temporal
La detección del movimiento es una de las tareas más importantes que se realiza en el cerebro. Gracias a ello, se puede adquirir información muy relevante de nuestro entorno. Por ejemplo,
permite agrupar objetos, estimar profundidades o estructuras tridimensionales.
14
Sistemas de Visión AER
Los sensores AER de contraste espacial se desarrollaron con posterioridad a los de contraste
temporal. El primer trabajo destacable en esta línea fue presentado por Mallik y sus colegas en
2005 [7]-[8]. Era un dispositivo basado en frames que incorporaba algunas de las ventajas de los
sistemas AER. Básicamente computaba los cambios relativos en iluminación entre frames consecutivos. Los eventos se generaban en instantes de tiempo discretos. El sensor combinaba las ventajas típicas de los sensores APS (Active Pixel Sensor) convencionales con algunas de los
sistemas AER.
Posteriormente, hay que destacar el trabajo de Lichsteiner y Delbrück [10]-[9]. Los autores
fueron capaces de implementar un sensor de contraste temporal puramente AER con algunas
características bastante destacables: alto ancho de banda (>2KHz), baja latencia ( 15µs @1Klux),
amplio rango dinámico (120dB), bajo consumo (24mW) y buena resolución (128 x 128 píxeles).
El segundo objetivo de la tesis consistió en implementar un sensor AER de contraste temporal que mejorara algunas de las prestaciones del sensor antes comentado. En particular, la nueva
retina fue diseñada para poder detectar movimiento a muy alta velocidad y con un tiempo de respuesta bastante bajo. Además, el tamaño del píxel se redujo con respecto a la retina de Lichsteiner
y sus colaboradores.
2 Corrientes de Polarización Programables (The I-Pot System)
2.1 Introducción
Los circuitos analógicos requieren en general de una serie de corrientes de polarización para
su funcionamiento. Dichas corrientes son normalmente proporcionadas por un circuito externo,
comúnmente situado en la periferia del chip y destinado a tal fin. Sin embargo, a veces, es deseable ajustar los valores de estas corrientes de polarización para poder compensar variaciones del
proceso y el mismatch de los componentes del circuito. Otras veces, cuando se testa un chip, simplemente se desea tener la posibilidad de experimentar con valores distintos de polarización. Para
ello, una solución trivial es disponer de un pin externo por cada corriente de polarización del circuito. Sin embargo, el número de pines disponibles es limitado.
Con la intención de generar bias totalmente programables y solventar la limitación antes
comentada, se ideó el I-Pot Estocástico [35]. Los I-Pots son fuentes de corrientes programables
mediante una palabra digital de control. Cada I-Pot es una celda que recibe una corriente de referencia a la entrada y proporciona la corriente deseada con precisión por debajo de los pico-amperios. El número de pines externos para programar, caracterizar y alimentar los I-Pots es tres,
independientemente del número de celdas programables que se incorporen al chip. En un proceso
de 0.35µm estándar, el área de un I-Pot es de 130µm × 68µm , lo que equivale a una tercera
parte del pad al que reemplaza.
2.2 Arquitectura de los I-Pots
Los I-Pots están basados en estructuras en escalera construidas con transistores MOS [32][28]. En la Figura 1, se muestra una estructura en escalera genérica construida con transistores
MOS y configurada en modo de fuente de corriente. Todos los transistores tienen un tamaño proporcional a uno unitario de dimensiones W/L. El factor de proporcionalidad puede ser 1, N-1, N/
(N-1). De esta forma, las corrientes de cada rama Ii tienen un factor de atenuación de valor N con
respecto a la rama previa. El I-Pot Estocástico usa dos de estas estructuras en escalera, tal como
podemos ver en la Figura 2, donde se presenta el diagrama de bloques completo de un I-Pod. La
15
Sistemas de Visión AER
I REF
N−1
N−1 M
a
I
I m = REF
m−1
I
I = REF
2 N
I =I REF
1
Ma
Ma
N
I REF
N
m−1
(N−1)
N−1
N−1
N−1
Ma
N
I m+1 =
1
Ma
Mb
N
Mf
Mb
N−1
N−1
Fig 1: Estructura en escalera genérica con N ramas. Se indica el valor de la corriente de
cada rama.
Range
Selector
Ladder (N=10)
IREF
IREFcopy
Current
Mirror
sw1
sw2
sw3
Irange
Io
VG
Ia
To external
I−test
(N=2)
Data−in
Stochastic
Ladder
wval
wrange
wsign wtest
Data−out
Clock
to j−th
bias
Fig 2: Diagrama de las distintas partes de cada I-Pot o celda programable.
primera de ellas (Range Selector Ladder) con un valor de atenuación de N=10 selecciona el rango
de operación. Tiene 6 ramas de salida y la corriente de referencia de entrada IREF puede tomar un
valor a la salida comprendido entre IREF e IREF /106. La segunda de ellas (Stochastic Ladder) con
un valor de atenuación de N=2 permite hacer combinaciones binarias de sus ramas para generar
un valor preciso de intensidad a su salida.
En nuestro diseño, la segunda estructura en escalera hace uso de transistores de tamaño
pequeño. De esta forma, se consiguen valores muy dispares que cubren de forma homogénea todo
16
Sistemas de Visión AER
b
b
15 14
b
13
b
12
b
b
11 10
b
3
b
2
b
0
b
1
I
out
I range
I
dump
W/L
W/2L
a
I1
W/2L
a
I2
W/L
W/2L
W/2L
a
I3
W/2L
W/2L
W/L
W/2L
W/2L
W/L
W/L
b
I1
a
I8
a
I7
b
I2
W/2L
b
I3
W/2L
b
I7
W/2L
b
I8
W/2L
W/L
W/L
Fig 3: Esquemático de una estructura en escalera con factor de atenuación N=2 y ramas de
salida duplicadas.
el rango de operación. Además las ramas de salida de esta escalera están duplicadas tal como
puede verse en la figura Figura 3. De esta forma se consigue un rango continuo de valores posibles de salida sin que haya grandes discontinuidades entre los valores de corriente que se pueden
obtener con cada I-Pot.
Volviendo al esquema completo de cada I-Pot de la Figura 2, vemos que cada celda de
corriente es alimentada por una corriente externa de referencia IREF a través de un espejo PMOS.
Dicha corriente es la entrada a la primera escalera que selecciona el rango de operación. La salida
de este bloque es la entrada de la escalera con rango de atenuación N=2 que permite hacer un
ajuste fino de la corriente de salida. Finalmente a la derecha, tenemos una circuitería que permite
seleccionar el signo de la corriente y decidir si la salida se conecta a un pin externo de test o bien
alimenta al circuito que corresponda. Todo se programa mediante una palabra digital de control.
Con wrange se selecciona el rango de operación, con wval se selecciona uno de los posibles valores
dentro del rango elegido, con wsign se elige el signo de la salida y con wtest se decide si la corriente
de salida va a un pin externo de test.
La única desventaja de usar los I-Pots es que cada uno necesita ser caracterizado una vez
que el chip ha sido fabricado. Sin embargo, el proceso es sencillo y sólo necesita realizarse una
vez. Para ello, sólo es necesario usar un ordenador que vaya cargando los registros con todos los
posibles valores y un amperímetro conectado al pin de salida. El procedimiento detallado para
caracterizar los I-Pots es el siguiente:
1.
2.
Cada I-Pot necesita ser caracterizado independientemente. Por tanto, sólo el bit wtest de uno
de los I-Pots puede estar activo. Así, sólo uno de ellos está conectado a la línea externa Itest.
Se barren los dos signos del I-Pot activo.
17
Sistemas de Visión AER
3.
4.
Para el I-Pot y el signo seleccionado, se barren todos los rangos de corriente mediante la palabra digital wrange.
Para el valor elegido de wrange, se barren las 20 ramas de salida, se miden cada uno de los
posibles valores de salida a través del pin externo Itest y se almacenan en el ordenador.
Una vez que se ha completado el proceso de medida del I-Pot, tendremos almacenados en el
ordenador un total de 2 signos x 6 rangos x 16 ramas=192 valores de corriente. Para cada signo y
para cada rango, podemos generar 216 combinaciones. Finalmente, con un sencillo programa de
ordenador podemos obtener los valores óptimos de wrange y wval que minimizan el error cuando se
desea obtener una determinada corriente de salida.
2.3 Resultados experimentales
Para caracterizar los I-Pots, se fabricaron un conjunto de ellos en tecnología AMS de
0.35µm . La escalera con factor de atenuación N=2 se implementó con 16 ramas de salida. El
tamaño de los transistores unitarios de esta escalera era W = 1µm y L = 0.7µm . Se utilizó una
I REF = 100µA .
Si obtenemos todos los posibles valores de salida de un I-Pot y los ordenamos, podemos
definir el incremento entre dos valores consecutivos, ∆ rel , como
In – In + 1
∆ rel = 2 ------------------------In + In + 1
(1)
Y conociendo el valor de ∆ rel , podemos expresar la precisión de cada valor de salida en
bits:
1⁄2
n bits
= ∆ rel ⇔ n bits = – log 2 ( ∆ rel ) = – ( ln ∆ rel ) ⁄ ( ln 2 )
(2)
En la Figura 4, se han representado los incrementos entre los valores medidos y ordenados
de intensidad para cada uno de los 6 rangos de operación. Puede verse que la mayor precisión se
alcanza en las zonas centrales de los rangos de operación, mientras que los peores resultados se
obtienen en los extremos de los rangos. Para el valor más pequeño del rango de operación, la
medidas presentan estriaciones. Ello puede deberse a errores debidos al instrumento de medida.
En la Figura 5, se presentan la precisión en bits de todos los valores posibles que pueden obtenerse con todos los rangos de operación. Vemos que hay zonas en las que se alcanza una resolución superior a los 13 bits. En las regiones correspondientes a los extremos de los rangos, se
alcanza una precisión de 8 bits. Obviando los extremos de los 6 rangos superpuestos, la precisión
mínima que se alcanza es de 8.1 bits aproximadamente.
2.4 Conclusiones sobre los I-Pots
Los I-Pots son una poderosa herramienta para polarizar y testar circuitos neuromórficos.
Permiten generar un número arbitrario de corrientes de polarización programables con una precisión por debajo de los pico-amperios. Tan sólo tres pines de salida son necesarios para incorporar-
18
Sistemas de Visión AER
14
14
14
12
12
12
10
10
10
8
8
8
6
6
6
4
4
4
2
2
2
0
0
0
1
2
3
0
1
2
0
3
−10
−8
14
12
12
12
10
10
10
8
8
8
6
6
6
4
4
4
2
2
2
0
0
4
4
x 10
14
2
2
x 10
14
0
0
−9
x 10
6
0
5
−7
0
10
0
5
10
−6
x 10
15
−5
x 10
x 10
Fig 4: Incrementos relativos entre valores consecutivos ordenados y expresados en bits. El
eje de ordenadas indica la precisión en bits, mientras que el de abcisas indica el valor de la
corriente.
13
12
11
Precision (bits)
10
9
8
7
6
5
4
−11
10
−10
10
−9
10
−8
10
−7
10
−6
10
−5
10
−4
10
Measured Current (A)
Fig 5: Incrementos relativos entre valores de corriente cuando se ordenan todos los valores
obtenidos en cada rango. El eje de abcisas indica el valor de la corriente y el eje de ordenadas indica la precisión en bits.
19
Sistemas de Visión AER
los en un chip. La única desventaja es que deben ser caracterizados una vez que se haya fabricado
el chip. El área de un I-Pot es de 130µm × 68µm .
Los I-Pots fueron testados por separado y usados para polarizar y testar las retinas que se
describirán a continuación en este documento. Han demostrado ser muy útiles en el test de circuitos neuromórficos donde se necesitan números elevados de corrientes de polarización y el número
de pines externos disponibles es limitado.
3 El Nuevo Método de Calibración (Newcalib)
3.1 Introducción
Durante los últimos 20 años, una gran variedad de sistemas VLSI neuromórficos han sido
publicados. Normalmente, tales sistemas están compuestos por una gran número de píxeles que
forman un array. La tendencia actual es reducir al máximo el tamaño de los píxeles y el consumo.
Para ello, corrientes del orden de nano amperios o menores tienen que ser usadas. Esto da lugar
necesariamente a un alto mismatch. A pesar de los sistemas VLSI neuromórficos, tienen un gran
potencial, aún no están suficientemente maduros para ser comercializados. Una de las razones de
que esto ocurra, es el alto mismatch que presentan. En concreto, una de las grandes limitaciones
de las retinas de contraste espacial implementadas hasta la fecha ha sido el mismatch que empeoraba la calidad de las imágenes captadas.
Si queremos reducir el mismatch sin aumentar el tamaño de los transistores, la única solución viable es la calibración. Existen técnicas de calibración para redes neuromórficas [14]-[27],
basadas en DACs compactos realizados con estructuras en escalera programables [28]. El problema de tales estructuras, era que la precisión estaba limitada a un punto (a un valor concreto
para el cual se calibraba). En nuestro caso, hemos diseñado un nuevo sistema de calibración multi
punto [13], mucho más compacto que una estructura en escalera, que permite escalar las corrientes, una vez que el sistema ha sido calibrado, sin que la precisión del calibrado se degrade de
forma notoria. Tal sistema fue testado e incorporado en los píxeles de la retina de contraste espacial para compensar su mismatch.
3.2 Nuevo Sistema de Calibrado (Newcalib)
El nuevo método de calibración [13] está basado en transistores MOS cuya longitud es programable mediante una palabra digital de control, wcal. En la Figura 6, podemos ver los esquemáticos de la nueva estructura de calibrado y su símbolo. El circuito de Figura 6(a) entre los
terminales D, G y S, se comporta exactamente igual que un transistor MOS digitalmente ajustable
mediante la palabra digital w cal = { b N – 1 b N – 2 …b 2 b 1 b 0 } Y su longitud equivalente viene dada
por L eq = ( W ⁄ 2 ) × g ( w cal ) , donde
N–1
g ( w cal ) =
bi
----------------∑ 2N – 1 – -i
i=0
20
(3)
Sistemas de Visión AER
VDD
D
bN−1
IREF
S N−1
G
Ical
W
L
b1
S1
W
L g( wcal )
b0
S0
(a)
W
L
wcal
S
Ib
(b)
Fig 6: (a) Esquemáticos del transistor MOS con longitud controlada digitalemente. (b)
Posible aplicación para calibrar una fuente de corriente.
La palabra digital es almacenada en una memoria estática de tipo RAM cuando el circuito
empieza a funcionar. La longitud equivalente del transistor puede ser ajustada entre 0 y
N–1
N–1
2 – (1 ⁄ 2
) en pasos de 1 ⁄ 2
. Este transistor puede ser usado como parte de un espejo
de corriente, tal como es mostrado en la Figura 6(b), para proporcionar una corriente de calibración I cal = I REF × ( g ( w cal ) + 1 ) . En la Figura 7, se muestran los resultados de una simulación
Montecalo con 100 iteraciones distintas del circuito de la Figura 6(b). Se utilizó una palabra digital de 4 bits. En el eje de abcisas, se fue variando el valor de la palabra de calibrado, wcal, y en el
eje de ordenadas, se representa el valor de la intensidad de calibrado resultante. El circuito se
polarizó con una corriente de valor I REF = 3nA y tenía transistores unitarios de tamaño 1 µ /
4 µ m. La simulación se realizó usando modelos de un proceso CMOS estándar de 0.35 µ m.
3.3 Uso de Circuitos Translineales para el Calibrado
El circuito de la Figura 6(b), tiene el grave inconveniente de que sólo permite calibrarlo en
un punto. Es decir, si escalamos la corriente de referencia I REF el calibrado no se conserva y la
corriente I cal no se escala en la misma proporción. Para conseguir esto, introducimos bucles
translineales. De esta forma, se consigue mantener fijas algunas corrientes, incluida la del transistor con longitud programable, y se escalan otras. Esto se muestra en la Figura 8. La circuiteria
encerrada dentro de las líneas discontinuas se replica dentro de cada píxel, mientras que el resto
aparece una vez en la periferia del chip. El grupo de transistores que va de M1 a M4 forma un
bucle translineal, lo cual implica que I oi = I 1 I 2 ⁄ I i3 . La corriente local I i3 se copia desde la
corriente periférica I 3 , mediante un espejo de corriente local con una un transistor de longitud
programable. Para conseguir un rango de calibración con factor 2, se incluyen dos transistores en
serie en el espejo de corriente de salida. Uno de tamaño fijo W ⁄ 2L y el otro de tamaño calibrable. Consecuentemente, I i3 = I 3 ⁄ ( 2 + g ( w cali ) ) y
21
Ical
Sistemas de Visión AER
wcal
Fig 7: Simulación Montecarlo (con 100 iteraciones) del circuito de la Figura 6(b), usando un
transistor MOS de longitud ajustable con una palabra digital de 4 bits.
wcal
i
I1
W
Lg(wcal )
i
M2
W
2L
W
L
M4
M1
M3
I2
Ii 3
Ioi
I3
Fig 8: Circuito translineal usado para permitir ajustar el rango del circuito de calibrado.
22
Sistemas de Visión AER
wcal
i
I1
M2
M
W1 m
L1
W1
L1g(wcal )
i
W1
L1g(wadj )
wadj
M1
M3
I2
Ii 3
M4
Ioi
I3
Fig 9: Primera estrategia para optimizar el rango de calibración.
I1 I2
I oi = --------- ( 2 + g ( w cal i ) )
I3
(4)
Con este circuito, uno puede mantener (después de calibrar) constante las corrientes I 3 (e
I i3 ) e I 1 , mientras I 2 se ajusta globalmente para escalar por arriba o por abajo las corrientes
locales I oi .
3.4 Estrategias para el Calibrado
Basándonos en las estructuras para el calibrado con bucles translineales, se propusieron y
testaron dos posibles esquemas de calibrado con bucles translineales. En la Figura 9 y la Figura
10, se pueden ver las dos estrategias usadas para implementar el calibrado.
Con la primera estrategia de calibrado, mostrada en la Figura 9, dos transistores con longitud
ajustable mediante una palabra digital son usados. Uno de ellos es ajustado localmente, como en
Figura 8, pero el otro es ajustado globalmente y su longitud se varía de la misma forma para todos
los píxeles del array. De esta forma, todas las puertas de los segmentos del transistor son controladas desde la periferia (véase Figura 9). Como resultado,
I1 I2
I oi = --------- ( g ( w adj ) + g ( w cal i ) )
I3
(5)
Con la segunda estrategia de calibrado, mostrada en la Figura 10. El bucle translineal se ha
replicado dos veces, de forma que se tienen dos bucles translineales en paralelo. Uno de ellos usa
la palabra de calibración local, w cal i . El otro es ajustado de forma global a todo el array y sólo el
transistor de salida M x es replicado una vez por píxel. Esto permite usar un transistor de salida con
un tamaño mayor y consecuentemente menos mismatch. El propósito del circuito translineal calibrable es compensar el mismatch de M x .
23
Sistemas de Visión AER
I’1
I’3
Mx
Ix
I’2
I1
wcal
i
M2
W1
L1
M4
M1
M3
Ii 3
I2
Ii 4
Ioi
I3
Fig 10: Segunda estrategia para optimizar el rango de calibración.
3.5 Resultados Experimentales
Para caracterizar los nuevos sistemas de calibrado, un chip prototipo se fabricó en un proceso CMOS de 0.35 µ m. Se fabricaron 20 DACs de corriente de 5 bits cada uno. Diez de ellos
usaban la estrategia de calibración de la Figura 9, y los otros diez usaban la estrategia de calibración de la Figura 10.
Cada uno de los 10 primeros DACs usaba 5 réplicas del circuito de la Figura 9, una por cada bit.
La corriente nominal de salida de cada uno ( I oi ) se ajustaba para que fuera escalada variando la
palabra digital. Consecuentemente, en la periferia, necesitamos 5 grupos de fuentes de corriente
{I 1 , I 2 , I 3 } y 5 grupos de transistores {M 1 , M 2 , M m }, uno por cada bit. Sin embargo, esos 5 grupos de fuentes de corriente periféricas y transistores son compartidos por todos los 10 DACs.
Cada uno de los segundos 10 DACs usa 5 réplicas del circuito de la Figura 10. De nuevo,
para cada uno de los 10 DACs, la circuitería se replicó 5 veces (una por cada bit), y la circuitería
periférica (fuera de las líneas con trazo discontinuo de la Figura 10) es compartida, por cada bit,
por los 10 DACs.
2
El área del layout del circuito dentro de la línea discontinua es 18 × 14µm para la Figura 9 y
2
17 × 14µm para el de la Figura 10.
La Figura 11(a) muestra las corrientes de salida medidas de forma experimental para las 10
réplicas del circuito de la Figura 9 cuando se fija w adj = 0 . Las corrientes periféricas de salida
24
Sistemas de Visión AER
4.5
4
5
(b)
(a)
4
3.5
3
3
nA
nA
2.5
2
2
1.5
1
1
0.5
0
0
5
10
15
20
wcal
25
0
0
30
5
10
15
wcal20
25
30
Fig 11: Corrientes de salida obtenidas experimentalmente para el circuito de la Figura 9
(a) para wadj=0, (b) para el valor óptimo de wadj. La línea horizontal en (b) es el valor objetivo, el cual es tocado o cortado por las 10 curvas.
eran I 1 = I 2 = I 3 = 10nA , y todas las palabras de calibrado w cali (i = 1,... 10) fueron barridas de
forma simultánea de 0 a 31. Después de repetir estas medidas para todos los valores posibles de
w adj , el valor óptimo para w adj corresponde con la situación donde el valor más alto de la
izquierda es más cercano al valor más bajo de la derecha. Esta situación se muestra en la Figura
11(b). En este punto podemos obtener las palabras de calibrado óptimas w cali que dan la mínima
variación. La máxima desviación de las corrientes de salida se obtiene bajo esas circunstancias es
∆I oi max = 0.57nA , lo que corresponde al 5.7%, de la corriente nominal I b = 10nA . Si ésta
fuera la fuente de corriente controlada por el bit más significativo de un DAC de corriente (con un
rango
máximo
de
20nA ),
esto
limitaría
la
precisión
del
DAC
a
– ln ( ∆I oi max ⁄ 2I b ) ⁄ ln 2 = 5.13bits . Para evaluar cómo la calibración se degrada cuando
cambian las condiciones de polarización, se barrió I 2 en la Figura 9 entre 100pA y 1µA . La
máxima desviación entre las 10 fuentes de corriente calibradas se muestra con un trazo con círculos en la Figura 12. El trazo con triángulos corresponde a medidas obtenidas antes de calibrar
( w cali = 0 , para todas la i). Podemos ver que las 10 muestras mantienen una precisión de 4 bits
para corrientes por encima de 3nA. El eje de abcisas es un promedio de I oi entre las 10 muestras.
También se muestra en la Figura 12 (trazo con cruces) la precisión que resulta después de calibrar
obtenida mediante simulaciones. Nótese que es muy optimista, excepto para el punto para el que
se calibró (10nA). La razón es que en este circuito en particular Figura 9 el calibrado se degrada es
debido al mismatch en el factor de forma del transistor. El mismatch de este parámetro no estaba
modelado en nuestro simulador.
De forma similar, la Figura 13 muestra la precisión medida antes y después de calibrar de 10
fuentes calibrables de corriente que siguen el esquema de calibración explicado en Figura 10.
Ahora el mismatch antes de calibrar era menor que en la Figura 12. Esto es porque ahora el área
usada por el transistor de longitud ajustable mediante una palabra digital M b en la Figura 9, está
disponible para el transistor M x de la Figura 10, con lo que puede ser de mayor tamaño. Con la
25
Sistemas de Visión AER
8
measured
simulated
uncalibrated
7
Precision (bits)
6
5
4
3
2
1
0
−10
10
−9
10
−8
10
Current (A)
−7
10
−6
10
Fig 12: Medidas de precisión de la fuente calibrable y configurable con el esquema propuesto en la Figura 9. Trazo con círculos: precisión medida tras el calibrado (con valor
óptimo de wcalj para cada una de las fuentes de calibrado). Las fuentes de corriente se calibraron a 10nA. Curva con triángulos: precisión medida antes de calibrar ( w cali = 0 para
todas las fuentes de corriente). Curva con cruces: precisión tras calibrar, obtenida con el
simulador.
estructura de la Figura 10 obtenemos mucha más precisión en el punto de calibrado (8.30 bits a
10nA), pero se degrada rápidamente, especialmente para corrientes altas. La precisión obtenida
mediante simulación, es ligeramente pesimista en el punto de calibrado (7.63 bits a 10nA), pero se
degrada de forma más optimista cuando las corrientes se alejan del punto de calibrado.
3.6 Conclusiones sobre el Sistema de Calibrado
Parte del trabajo realizado durante la tesis consistió en implementar un sistema de calibrado
compacto y eficiente que pueda ser empleado en sistemas neuromórficos. El sistema de calibrado
se testó de forma individual con un chip específico para ello. Existen dos estrategias para implementar el calibrado. Como posteriormente explicaremos, la retina de contraste temporal hace uso
de la segunda estrategia de calibrado (Figura 10).
4 La Retina AER de Contraste Espacial
4.1 Introducción
Las retinas están entre los primeros bloques AER que fueron publicados [1]-[2]. En concreto, las retinas de contraste espacial son de especial interés porque proporcionan un flujo de
26
Sistemas de Visión AER
9
measured
simulated
uncalibrated
8
Precision (bits)
7
6
5
4
3
2
1 −10
10
−9
10
−8
10
Current (A)
−7
10
−6
10
Fig 13: Medidas de precisión de la fuente calibrable y configurable con el esquema propuesto en la Figura 10. Trazo con círculos: precisión medida tras el calibrado. Las fuentes
de corriente se calibraron a 10nA. Curva con triángulos: precisión medida antes de calibrar ( w cal i = 0 para todas las fuentes de corriente). Curva con cruces: precisión tras calibrar, obtenida con el simulador.
datos comprimido, preservando la información necesaria para reconocer objetos. Este tipo de sensores pueden codificar la información de dos formas: La primera de ellas es generando una
secuencia de pulsos con una frecuencia proporcional al valor de la señal que se desea transmitir
(como luminancia o contraste). La segunda está basada en el modo TFS (Time to First Spike). En
este modo de funcionamiento, existe un reset global tras el cual los píxeles que detectan contraste
emiten un único pulso. La información se codifica como el tiempo transcurrido entre el reset y el
tiempo que tarda cada píxel en generar un pulso. Lógicamente los píxeles que detecten más contraste dispararán primero. A pesar de que el modo TFS está sujeto a las restricciones de los sistemas basados en frames, es un mecanismo altamente eficiente para comprimir la información.
Además, el tiempo de frame puede ser ajustado de forma dinámica mediante las etapas de posteriores de procesamiento.
Las salidas de las retinas de contraste espacial publicadas previamente [43]-[4]-[14] son normalmente una corriente de contraste por píxel I cont ( x, y ) . Dicha salida se computa como el
cociente entre la foto corriente obtenida al sensar la intensidad luminosa I ph ( x, y ) y el promedio
espacial de la foto corriente en una determinada región alrededor del píxel. El promedio se
obtiene normalmente mediante algún tipo de red difusiva
27
Sistemas de Visión AER
I ph ( x, y )
I cont ( x, y ) = I ref ---------------------I avg ( x, y )
(6)
donde I ref es una corriente global que escala el resultado. La intensidad de salida es siempre positiva (salida unipolar sin signo) y en el caso de que no haya contraste e I ph ( x, y ) = I avg ( x, y ) , la
intensidad de salida de cada píxel será igual a I ref . Esto supone una importante limitación porque
los píxeles consumirán ancho de banda de forma innecesaria. Además contradice o viola las ventajas de los sistemas AER, donde sólo la información relevante es transmitida.
En otro trabajo previo [14], el contraste se computaba mediante la fórmula de Weber,
I ph ( x, y )
I cont ( x, y ) = I ref  ---------------------- – 1
I avg ( x, y )
(7)
lo cual solventaba el problema del consumo de ancho de banda en ausencia de contraste y además
permitía generar una salida bipolar y detectar el signo del contraste. Para ello, una etapa de postprocesamiento, que anulaba los eventos generados en ausencia de contraste, era usada. Sin
embargo, esto hacía que la velocidad de respuesta de la retina se redujera un factor 10 aproximadamente.
En esta sección, vamos a presentar una nueva retina AER sensible al contraste espacial [23],
en la que el contraste se computa de la misma forma que en la eq. (7). El diseño se basa en el propuesto por Boahen [43], el cual ha sido mejorado para solucionar algunas limitaciones como el
elevado mismatch, la dependencia de las salidas con el nivel de luminosidad ambiente, y la dificultad de ajuste del dispositivo. También añade nuevas prestaciones. La salida es bipolar. Además, un mecanismo de umbralización ajustable se ha incorporado, de forma que los píxeles no
generan eventos, a menos que su salida supere un cierto umbral. Por último, la retina también
incluye un mecanismo global de reset que permite operar en modo TFS (Time to First Spike) con
independencia del nivel de luminosidad ambiente. Un prototipo de 32 x 32 píxeles en tecnología
CMOS de 0.35 µm fue fabricado y testado.
4.2 Píxel de Contraste Espacial de Boahen
El sensor de contraste espacial descrito en esta tesis solventa todas las limitaciones encontradas en el diseño de Boahen [44]. El comportamiento del circuito mostrado en la Figura 14 queda
aproximadamente descrito mediante las siguientes ecuaciones:
2
I h ( x, y ) = I ph ( x, y ) + a∇ I c ( x, y )
2
I c ( x, y ) = I u – b∇ I h ( x, y )
(8)
Obtener una expresión para la salida de cada píxel a partir de las ecuaciones anteriores no es
fácil ni intuitivo. La corriente de salida de cada píxel depende de todos los píxeles de la retina.
28
Sistemas de Visión AER
Ih
Mh
Vhh
Mc
Vcc
C
Ma
Ih
Iu
Mh
H
Mb Vhh
Mc
Iph
C
Ic
Vcc
Ma
Iu
H
Mb
Iph
Ic
Fig 14: Circuito original de Boahen para el cómputo del contraste espacial.
Una aproximación propuesta por Boahen para la salida de cada píxel [44] a partir de las ecuaciones (8) es
I ph ( x, y )
I cont ( x, y ) = I ref --------------------------------------〈 I ph〉 + I ph ( x, y )
(9)
donde 〈 I ph〉 es el promedio de la foto corriente en el vecindario próximo al píxel. Obviamente,
los píxeles más cercanos tendrán mayor influencia en la salida. Ajustando Icc, el alcance de los
píxeles que afectan a la salida puede ser controlado.
El circuito original de Boahen tiene una serie de limitaciones. En primer lugar, el mismatch
era comparable a la propia señal de interés de salida. En segundo lugar, la señal de salida se cambiaba para el mismo estímulo cuando se variaban las condiciones de iluminación. Tercero, los voltajes de control V cc y V hh en la Figura 14 tenían un rango de ajuste muy estrecho y crítico. En la
nueva versión de la retina, todos estos inconvenientes han sido corregidos.
4.3 Nuevo Diseño del Píxel
La Figura 15 muestra un diagrama de bloques de las distintas partes que componen el píxel,
indicando cuáles son las señales intercambiadas entre los distintos bloques. El píxel contiene tres
partes principales: (1) la parte de foto sensado y cálculo del contraste, incluyendo calibración, la
cual proporciona la corriente de salida que computa el contraste I cont independientemente del
nivel de iluminación; (2) el bloque de integración y disparo, que incluye la circuitería para ajustar
el periodo refractario, umbralización, y el modo TFS de disparo; (3) y la circuitería de comunicación del píxel que manda los eventos hacia la periferia e interacciona con ella. Una explicación
detallada de todos los elementos del píxel puede encontrarse en [23]. A continuación, vamos a
describir en detalle el primero de los bloques.
La Figura 16 muestra cómo el circuito de Boahen para el cómputo del contraste ha sido
modificado para incluir un esquema de polarización basado en corrientes, para controlar las seña-
29
Sistemas de Visión AER
Vrfr
Mh
Icont
Mc
Iph
i&f
Ical
calibration
Photo sensing and
contrast computation
to event read out
_
col +
col
TFS
pulse+
ev_rst +
_
pulse
_
ev_rst
+
−
Vth
Vth
integrate−and−fire
with thresholding
TFS reset and
refractory time
+ event block
Rqst
_
Ack
event block
to row
arbiter
AER pixel
communication
Fig 15: Diagrama de bloques con las partes principales que forman el píxel.
Ih
Mh
Ihh
Mc
Ma
H
Mb I
hh
C
Iph
Icc
I’u
Ih
Iu
Mh
Mc
Ic
Icc
Ma
C
Iu
H
Mb
Iph
Ic
Icont
Vb
wcal
Ical
Fig 16: Detalle del bloque de sensado de la foto corriente y computación del contraste.
les originales de control de las redes difusivas V cc y V hh , que pueden verse en el circuito original de la Figura 14. De esta forma, los voltajes en las puertas V cc y V hh tienden a seguir las
excursiones de tensión en los nodos ‘C’ y ‘H’.
La primera ventaja de usar este esquema es que las tensiones de los nodos del circuito se
adaptarán a las condiciones de iluminación ambiente. Por ejemplo, si todas las corrientes de los
30
Sistemas de Visión AER
foto diodos se escalan por arriba o por debajo por el mismo factor, el voltaje de todos los nodos
‘H’ seguirá los cambios de forma logarítmica. Como I u es constante, el voltaje del nodo ‘C’
seguirá, por tanto, el mismo desplazamiento en tensión. Como las corrientes de polarización I hh
e I cc son constantes y no varían, los voltajes en la puertas de los transistores M h y M c seguirán
la misma variación global de voltaje, adaptándose ellos mismos a los cambios globales de iluminación.
La segunda ventaja de este esquema es que atenúa el mismatch. Después de hacer un estudio
concienzudo del mismatch e identificar las principales fuentes de mismatch del circuito, se descubrió que el transistor M a y la corriente I u eran las fuentes dominantes. Esto puede ser entendido si
tenemos en cuenta lo siguiente. Las variaciones en la corriente I u afectan directamente a la componente en DC de I c , que será calibrada con I cal . El mismatch de M b es menos crítico porque
las variaciones de tensión en el nodo ‘C’ afectan a la red difusiva inferior y al cómputo de la
corriente promedio I h de la eq. (8). Por tanto, el efecto de su variabilidad es atenuado al hacer el
cómputo de la corriente promedio. Sin embargo, el mismatch de M a (las variaciones de V gs de
M a ) cambian directamente el voltaje en la fuente de M b , afectando directamente a la ganancia
del contraste de salida (el coeficiente ‘b’ en la eq. (8)), cuyo efecto no es directamente calibrado
por I cal . Consecuentemente, M a necesita ser dimensionado para minimizar el mismatch. El
efecto de I u es compensado mediante calibración, y el efecto de M a será atenuado por el nuevo
esquema de polarización. Nótese, que el mismatch en todos los transistores M a introducirá variaciones aleatorias en los nodos ‘H’ y ‘C’. Tales variaciones se transformarán en corrientes laterales
aleatorias a través de los transistores M h y M c . Las corrientes aleatorias a través de M h se añadirán a la corriente de salida I c y pueden ser compensadas mediante calibración. Sin embargo,
las corrientes aleatorias a través de los transistores M c afectan si fueron generadas por los foto
diodos. Gracias al nuevo esquema de polarización, un incremento en ‘C’ incrementará el voltaje
de la puerta del nuevo transistor NMOS, incrementando su voltaje en la fuente, y a su vez, incrementando el voltaje en la puerta de M c , lo cual reduce la corriente lateral. Un efecto similar ocurrirá con los transistores M h .
Finalmente, la tercera ventaja es un mecanismo más robusto de polarización de los transistores laterales. En el esquema original, las tensiones V cc y V hh tenían un rango de ajuste muy
pequeño y crítico (aproximadamente 100mV o menos). Ahora, las corrientes I cc e I hh pueden
ser variadas varias décadas, percibiendo aún sus efectos.
4.4 Resultados Experimentales
Se fabricó un prototipo de 32 x 32 píxeles de la retina AER de contraste espacial con signo.
El sensor se diseño y fabricó en una tecnología CMOS de 0.35µm con 4 metales y dos tipos de
polisilicio, con una alimentación de V DD = 3,3V . La Tabla 1 resume las especificaciones del
chip. La Figura 17 muestra una micro-fotografía del dado, de tamaño 2.5 x 2.6 mm2. El chip completo, excepto el anillo de pads, está cubierto con la capa de metal más alta dejando aperturas para
que los foto diodos puedan captar la luz. La Figura 17 también muestra el layout de un píxel indi-
31
Sistemas de Visión AER
Tabla 1: Especificaciones del chip.
tecnología
CMOS 0.35µm 4M 2P
tensión de alimentación
3.3V
tamaño del chip
2.5 x 2.6 mm2
tamaño del array
32 x 32
tamaño del píxel
80 x 80 µm2
fill factor
2.0%
eficiencia cuántica del
0.34 @ 450nm
fotodiodo
complejidad del píxel 131 transistores + 2 caps
consumo de corriente
65µA @ 10keps
vidual resaltando sus principales partes. El layout de cada píxel es simétrico y está especulado con
respecto al de sus vecinos más próximos. De esta forma, las líneas digitales ruidosas son compartidas entre los vecinos, al mismo tiempo que las líneas de alimentación, y las líneas sensibles al
ruido. También, las líneas más sensible a las perturbaciones son separadas de las ruidosas. El área
del píxel es 80 × 80µm 2 , incluyendo el rutado. El píxel estaba compuesto de 131 transistores y
2 capacidades (la capacidad del circuito de integración y disparo y la capacidad del circuito que
genera el reset y controla el período refractario).
La retina se sometió a múltiples experimentos y pruebas en condiciones muy diversas de iluminación para evaluar sus prestaciones. La Tabla 2 resume sus características más significativas.
Para una explicación detallada de cómo se tomaron las medidas, véase [23]. Las características
más significativas del sensor son: muy bajo FPN ( ∼ 0 .6% ), bajo consumo (0.66-66mW), baja
Tabla 2: Especificaciones del sensor de contraste espacial
Funcionalidad
Contraste
espacial
a
número
de eventos
Dependencia con el
nivel de
luminsidad
Latencia
NO
0.1-10ms
Rango
dinámico
FPN
>100dB
0.6%
32
Fill
Factor
2%
Tamaño
del píxel
µm
2
80x80
Proceso de
fabricación
Consumo
0.35 µm
4M 2P
0.6666mW
Sistemas de Visión AER
2.6mm
32x32 pixel array
Row Arbiter
2.5mm
Bias Generator
AER-out
Configuration Registers
digital
comm.
i&f
80µm
thresholding
photodiode
calibration
contrast
computation
Fig 17: Micro fotografía del dado de 2.5mm x 2.6mm, y zoom del layout del píxel de 80µm
x 80µm , indicando la localización de sus partes principales.
33
Sistemas de Visión AER
Fig 18:. Elementos naturales. De izquierda a derecha: tornillo, clip, ojo y cara de una niña.
Fig 19: Capturas de un clip en modo TFS con distinto número de eventos capturados, M.
latencia (0.1-10ms) y elevado rango dinámico (>100dB). En la Figura 18, pueden verse imágenes
captadas con la retina. Pese a la baja resolución, muestran el correcto funcionamiento del sensor.
Por último, la retina incorpora un modo de funcionamiento TFS opcional. En la Figura 19, se
muestra una imagen captada empleando este modo de operación. Se muestra la imagen capturada
para distintos números de eventos capturados tras el reset global. Puede apreciarse que con un
número de eventos, M=250, es posible distinguir la imagen original. El modo TFS es también útil
para eliminar el ruido. Nótese que los píxeles con información relevante disparan primero. Si se
escoge la duración del frame de forma adecuada, los píxeles que simplemente detectan ruido, no
dispararán nunca.
4.5 Conclusiones sobre la Retina AER de Contraste Espacial
Parte del trabajo de la tesis consitió en testar una retina AER de contraste espacial. En esta
sección se han descrito la nueva arquitectura del chip y se han resumido los resultados experimentales que se obtuvieron al caracterizar el sensor. La nueva retina hace uso del esquema de calibrado explicado en la Figura 9, obteniendo unos niveles de ruido bastante bajos en comparación
con el sensor previo propuesto por Boahen [44]. Otras ventajas y mejoras de la nueva implementación han sido explicadas en este apartado de la tesis.
34
Sistemas de Visión AER
5 La Retina AER de Contraste Temporal
5.1 Introducción
La detección de movimiento es una de las tareas más importantes que el cerebro humano
puede realizar. Gracias a ello, se puede conseguir mucha información relevante sobre nuestro
entorno. La detección del contraste posibilita el reconocimiento de estructuras tridimensionales y
permite agrupar los objetos que podemos ver. Tal es la importancia de la detección del contraste
espacial, que en el cerebro humano hay un canal especializado en ello.
Recientemente, varios sensores específicos de contraste temporal han sido publicados. El
primer sensor específico de contraste fue presentado por Mallik y sus colegas [7] en 2005 y estaba
basado en el trabajo previo de V. Gruev [48] que describía un sensor óptico basado en las variaciones temporales entre frames consecutivos. La retina de Mallik modificaba el tradicional pixel
activo (APS) para detectar un cambio absoluto y cuantizado en iluminación. Era un dispositivo
basado en frames con algunas de las ventajas de los sistemas AER. Recientemente, varios sensores de contraste espacial han sido publicados [9]-[24]. No son dispositivos basados en frames y
tienen varias ventajas sobre los sistemas convencionales. Sus características más importantes son
alto rango dinámico, capacidad de detectar objetos a muy alta velocidad y muy bajo consumo de
potencia.
Las retinas AER son inherentemente más rápidas que los dispositivos basados en frames. La
principal razón es que pueden detectar los cambios en iluminación de forma continua a lo largo
del tiempo y no hay tiempos muertos entre frames donde los cambios que ocurran no pueden ser
detectados.
Parte del trabajo de la tesis consistió en la implementación de un sensor AER de contraste
espacial. Está basado en el trabajo previo de Lichsteiner y sus colegas [9]. Sus píxeles responden
de forma asíncrona a los cambios relativos de intensidad. El sensor directamente codifica los cambios de luminosidad en la escena reduciendo la redundancia en los datos y el consumo de ancho
de banda. La nueva retina tiene características mejoradas y píxeles con menor tamaño. Como discutiremos más adelante, es especialmente adecuada para aplicaciones de seguimiento a alta velocidad y vigilancia.
5.2 Arquitectura del Píxel y Layout
La Figura 20 muestra el diagrama de bloques de las distintas partes del píxel. Hay cuatro
bloques principales: 1) El fotoreceptor, 2) Dos amplificadores para incrementar la sensibilidad y
disminuir el tiempo de respuesta, 3) Un integrador para detectar las variaciones temporales en iluminación, 4) Bloque de comunicación con la periferia.
Los dos amplificadores están conectados a un circuito externo que controla la ganancia de
forma automática (AGC). Éste se encarga de fijar la ganancia de los amplificadores a un valor
constante que no depende de la iluminación.
A continuación, vamos a describir cada uno de los bloques que forman el píxel y el circuito
externo que controla la ganancia de los amplificadores.
35
Sistemas de Visión AER
Amplificadores
Fotoreceptor
Vbp1
Mp
Mbp1
G1
Vout
McasVbn1
A1
G2
Vbn1
A2
Vcap
Interfaz de comunicación
Vrfr
Vref
Vg
Vcas
Circuito diferenciador
Ibn
5C
Vdiff
Rqst
Communica- Ack
tion Block
C
Vth+
Mm
Vldc1
Vldc2
Vth-
Reset
Fig 20: Diagrama de bloques de las distintas partes que componen el píxel.
E.
Fotoreceptor
La primera etapa es el fotoreceptor. Este circuito responde logaritmicamente a los cambios
en la iluminación local. Hay un offset en la etapa de salida, Vout, que será amplificador en la
siguiente etapa de amplificación. Esto no es un problema porque será eliminado con el circuito
diferenciador, el cual sólo responde a variaciones en la iluminación local. Ajustando, Vg, Vbp1 y
Vcas, podemos variar la velocidad de respuesta del fotoreceptor. El voltaje a la salida viene dado
por la siguiente expresión,
I ph
I ph
V out = κ p V g + nU T ln  ------ = K + nU T ln  ------
I sp
I sp
F.
(10)
Amplificadores
Para incrementar la sensibilidad y el tiempo de respuesta de la retina, dos amplificadores se
añadieron después del fotoreceptor. Ambos trabajan en la región de saturación. De esta forma, la
ganancia total de cada píxel puede ser incrementada con respecto a la del diseño previo propuesto
por Lichsteiner y sus colaboradores [9]. En ese diseño, la ganancia total del píxel era igual al
cociente entre las dos capacidades del circuito derivador. Para conseguir una ganancia total de 20
con un buen macheo entre píxeles, grandes capacidades eran necesarias. Si usamos amplificadores, podemos decrementar la ganancia del circuito diferenciador. Esto significa que podemos
36
Sistemas de Visión AER
Vin
M1
M2
Vout
V1
M3
Ibn1
Vldc
Fig 21: Esquemático de los amplificadores.
reducir el tamaño de las capacidades y conseguir así mejor fill factor y mayor ganancia al mismo
tiempo. Mayor ganancia implica mayor velocidad y mayor sensibilidad al contraste. Para aplicaciones de alta velocidad, alta ganancia y sensibilidad son deseables porque los cambios en la iluminación local ocurren muy rápido. Como discutiremos con posterioridad, alta ganancia también
implica bajo latencia. El único inconveniente de usar amplificadores es un aumento del consumo.
Si comparamos el nuevo diseño con el de Lichsteiner [9], la total ganancia ha sido incrementada y el área del píxel se redujo a la mitad.
La Figura 21 muestra los esquemáticos de un amplificador. Los transistores M2 y M3 trabajan en inversión fuerte. El valor de Vldc es fijado por el circuito externo de control de la ganancia.
De esta forma, los niveles de DC a la salida no dependen de la iluminación. M1 es un seguidor de
fuente que introduce un desplazamiento en el voltaje en Vin que puede ser ajustado variando Vbn1,
por tanto V 1 = V in – ∆ V . Si M2 y M3 están en inversión fuerte,
W
W
------2- β n2 ( V 1 – V ldc – V Th ) 2 =  ------3- β p3 ( V dd – V out – V Tp ) 2
 L2 
 L3 
(11)
Resolviendo para Vout, y sabiendo que V 1 = V in – ∆ V ,
( W 4 ⁄ L 4 )β 4
- ( V – ∆V – V Tp – V ldc )
V out = V dd – V Th – --------------------------( W 3 ⁄ L 3 )β 3 in
(12)
Donde ∆V se fija a un valor constante. Un valor recomendable es ∆V ≈ 0.7V . En cada píxel, hay
dos amplificadores (A1 y A2), y la ganancia total a la salida de la etapa de amplificación es
GT = G1 G2 =
( W 2 1 ⁄ L 2 )β 21 ( W 22 ⁄ L 2 )β 22
1
2
-------------------------------- -------------------------------( W 3 1 ⁄ L 3 )β 31 ( W 32 ⁄ L 3 )β 32
1
37
2
(13)
Sistemas de Visión AER
En nuestro caso particular, ambos amplificadores se diseñaron para tener la misma ganancia
G 1 ≈ G 2 y ∆V cap = G 1 G 2 ∆V out . La componente DC a la salida es eliminada por el circuito
diferenciador. Simplemente hemos de preocuparnos de que la ganancia de los amplificadores no
sature la salida de alguno de ellos. El mismatch en la ganancia de los amplificadores es relativamente bajo (comparable al mismatch en los comparadores de la etapa de integración y disparo) y
no supone una limitación.
G.
Circuito derivador
Esta etapa responde a cambios relativos en el voltaje de la señal de entrada con una ganancia
igual a G 3 = 5C ⁄ C . El mismatch de las capacidades es pequeño. El circuito diferenciador también elimina la componente de DC en la señal de entrada. Podemos expresar las variaciones de
voltaje a la salida como
∆V diff = – G 3 ∆V cap = – G 1 G 2 G 3 ∆V out
(14)
I ph ( ∆t + t )
∆V out = nU T ln  -------------------------
 I ph ( t ) 
(15)
Y de la ecuación (10),
La ganancia total a la salida de esta etapa es G t = G 1 G 2 G 3 .
H.
Circuito de comunicación con la periferia
La Figura 22(a) muestra los esquemáticos del circuito de comunicación con la periferia. Los
píxeles pueden detectar el signo del contraste temporal. Para ello, hay dos comparadores con diferentes umbrales ajustables Vhigh y Vlow que generan eventos positivos o negativos (contraste con
signo) cuando Vdiff está por encima del umbral positivo o por debajo del negativo, respectivamente. También se incorporó un circuito refractario (véase Figura 22(b)) que permite controlar el
tiempo entre eventos consecutivos que un píxel individual puede generar. Esto es especialmente
útil cuando hay píxel con alta actividad y se desea controlar la actividad máxima en el bus AER.
Finalmente, hay dos bloques idénticos de comunicación con la periferia, tal como se ve en la
Figura 22(c), los circuitos de comunicación están inspirados en las técnicas de Boahen para mandar eventos al exterior mediante un bus paralelo [45]. Cuando se generan eventos con signo, cada
píxel necesita proporcionar dos señales por columna y por signo del evento, col+ y col-. Este concepto ya fue testado e implementado en la retina de contraste espacial [23] y diseños previos [27]
que generaban eventos con signo.
5.3 Circuito Externo de Control de la Ganancia
Este bloque está situado en la periferia del chip. En la Figura 23(a), podemos ver los esquemáticos se la circuitería de control de la ganancia de los amplificadores. Sus funciones son mantener constantes los niveles de DC a la salida de los amplificadores de la Figura 21, así como su
38
Sistemas de Visión AER
col+
(a)
Vhigh
col-
pulse+
Vrfr
Vref
ev_rst-
event block
Ack
Refractory
Circuit
ev_rst+ ev_rst-
Vlow
pulse-
(c)
Vdiff
(b)
Rqst
ev_rst+
Vdiff
Crfr
- event block
col
Vrfr
Rqst
ev_rst-
pulse
ev_rst+
ev_rst
Ack
Vref
Fig 22: (a) Neurona de integración y disparo con módulo AER de intercomunicación con
la periferia. (b) Detalle del circuito refractario. (c) Detalle del bloque AER de comunicación con la periferia dentro del píxel.
ganancia, independientemente de los valores de luminancia del chip. Esto se consigue ajustando
los voltajes Vldc1 y Vldc2 cuando varía la iluminación del chip. El bloque de control de la ganancia
tiene amplificadores idénticos a los de la Figura 21. Las salidas de esos amplificadores, A1 y A2,
son fijadas a unos valores constantes de tensión que podemos ajustar, Vodc1 y Vodc2. Para ello, las
salidas de los amplificadores, A1 y A2, están conectadas a dos amplificadores operacionales con
39
Sistemas de Visión AER
Vblc2
Vblc1
(a)
Vbpn1
A2
A1
Vg
Vbn1
Vcas
Vodc1
Vbn1
Vodc2
I ph
C1
C3
C2
(b)
Vbrf
Ibrf
Vg
Vg
I ph
Vcas
Fotoreceptor
Ilcm
I ph1
I ph 128
Foto diodos
gnd01
Vdd01
Espejo de corriente
sub-pico-ampérico
Fig 23: (a) Bloque de control automático de la ganancia. (b) Detalle de la circuitería usada
para sensar la iluminación promedio del chip. Se han señalado sus partes principales.
40
Sistemas de Visión AER
Tabla 3: Especificaciones del Chip.
tecnología
CMOS 0.35µm 4M 2P
tensión de alimentación
3.3V
tamaño del chip
5.54 x 5.69 mm2
tamaño del array
128 x 128
tamaño del píxel
35.7 x 35.5 µm2
fill factor
8%
eficiencia cuántica del
0.34 @ 450nm
fotodiodo
complejidad del píxel
58 transistors + 3 caps
consumo de corriente
44mA @ 10keps
realimentación negativa que se encargan de fijar los valores de tensión. De esta forma, los valores
de Vldc1 y Vldc2 serán idénticos en la etapa de amplificación y el bloque de control de la ganancia.
La corriente de entrada al bloque AGC es la foto corriente promedio sensada por 128 foto diodos
situados alrededor de la matriz de píxeles. En la Figura 23(b) podemos ver el circuito que sensa el
valor promedio de la foto corriente. Tiene 128 foto diodos y un fotoreceptor idéntico al de la
Figura 20. Cada uno está conectado a una capacidad para hacer que el circuito de control de la
ganancia sea insensible a cambios muy rápidos en los niveles de luminosidad. Este circuito sólo
debe detectar cambios lentos en la iluminación promedio del chip. El valor de esta capacidad fue
elegido para que la constante de tiempo fuera del orden de los milisegundos. Este tiempo es suficiente para adaptarnos a un cambio global en la iluminación. Por ejemplo, cuando salimos fuera
de una habitación y entramos en una zona muy iluminada con luz solar. Un tiempo de respuesta
excesivamente bajo, podría dar lugar a problemas de inestabilidad en el circuito. Finalmente, hay
un espejo sub-pico-ampérico [36] que copia la foto corriente sensada.
5.4 Resultados experimentales
Se fabricó un sensor AER de contraste temporal (con signo) de 128 x 128 píxeles. La retina
fue diseñada y fabricada en una tecnología CMOS de 0.35 µ m con cuatro metales y doble polisilicio con una tensión de alimentación V DD = 3.3 V. La Tabla 3 resume las especificaciones del
chip. La Figura 24 muestra una micro-fotografía del dado, de tamaño 5.689x5.538mm2. El chip
completo, excepto el anillo de pads, está cubierto con la capa de metal más alta, dejando aberturas
encima de los foto diodos. La Figura 24 muestra el layout de un píxel individual resaltando sus
partes principales. El layout de cada píxel es una especulación simétrica de los píxeles de su alrededor. De esta forma, las líneas digitales ruidosas son compartidas entre vecinos, al mismo tiempo
que las líneas de alimentación y las líneas sensibles al ruido. Al mismo tiempo, las líneas sensi2
bles al ruido se separan de las ruidosas. El área del píxel es 35.5x35.7 µm , incluyendo el rutado.
El píxel está formado por 58 transistores y 3 capacidades (las dos del divisor capacitivo y la del
circuito refractario).
41
Sistemas de Visión AER
5.538mm
bias
generators
5.689mm
128x128 pixel array
AER out
Global
AGC
AER communication block
refractory
capacitor
capacitive divider
35.5 µ m
photo diode
amplifiers
comparator
35.7 µ m
Fig 24: Micro-fotografía del dado 5.538mm x 5.689mm, y detalle del píxel de 35.7 µ m x
35.5 µ m, indicando la localización de sus partes principales.
42
Sistemas de Visión AER
Tabla 4.: Especificaciones del Sensor de Contraste Temporal.
Funcionalidad
Contraste
Temporal a
número
de eventos
Dependencia con el
nivel de
luminsidad
Latencia/
Ancho de
banda
NO
7.8Keps in 32.8ms
3.6 µs /
8KHz
Rango
Dinámico
FPN
>100dB
2.8%
20Keps in 30ms
Fill
Factor
8%
Tamaño
del píxel
µm
2
35x35
Proceso de
Fabricación
Potencia
0.35 µm
4M 2P
132185mW
1.8Keps in 32.8ms
Fig 25: Algunas capturas de escenas naturales. De izquierda a derecha: Mano, cara y
hombre. La objetivo usado era de 16mm.
La retina se sometió a múltiples experimentos y pruebas en condiciones muy diversas de iluminación para evaluar sus prestaciones. La Tabla 4 resume sus características más significativas.
Entre éstas cabe destacar las siguientes: muy baja latencia ( 3.6µs @25Klux), buen fill factor
(8%), elevado ancho de banda (>8KHz) y elevado rango dinámico (>100dB). En la Figura 25 ,
pueden verse imágenes captadas con la retina.
5.5 Conclusiones sobre la Retina AER de Contraste Temporal
Parte del trabajo de la tesis consitió en testar una retina AER de contraste temporal. En esta
sección se han descrito la nueva arquitectura del chip y se han resumido los resultados experimentales que se obtuvieron al caracterizar el sensor. La retina hace uso de una etapa de amplificación
tras el fotoreceptor logarítmico. De esta forma, se aumenta la sensibilidad, el ancho de banda y se
reduce la latencia y el tamaño del píxel, con respecto al diseño previo propuesto por Lichsteiner y
sus colegas [9].
43
Sistemas de Visión AER
6 Conclusiones y Líneas Futuras de Investigación
En el presente documento se han descrito dos sensores AER distintos capaz de detectar el
contraste espacial y temporal, además del signo. El primero de ellos es un sensor de contraste
espacial que solventa las limitaciones de otros sensores del mismo tipo publicados previamente.
Dicho sensor, está basado en el trabajo previo de Boahen [21]-[22] y resuelve alguna de sus limitaciones. Se elimina el consumo innecesario de ancho de banda cuando no hay contraste, proporciona salida unipolar, el cómputo del contraste no depende de la iluminación del chip, se reduce el
consumo de potencia y el FPN se ha reducido notablemente. El mismatch era una de las limitaciones que ha afectado tradicionalmente a los sistemas neuromórficos y en particular a las retinas de
contraste espacial basadas en el cómputo del contraste mediante el cociente de intensidades. Por
esta razón, se desarrolló un nuevo sistemas de calibrado compacto para redes neuromórficas [13].
El nuevo sistema tiene un tamaño menor que otros implementados previamente [14]-[27]. Además, permite escalar las corrientes de calibrado sin que el calibrado se degrade. El circuito de calibrado se probó de forma individual y los resultados obtenidos fueron satisfactorios.
Posteriormente, se incorporó a la retina de contraste espacial para solventar los problemas de mismatch, reduciendo el valor del FPN de forma notable.
El segundo de los sensores presentados en la tesis es un sensor de contraste temporal. Esta
retina es especialmente útil para el seguimiento de objetos a alta velocidad y aplicaciones de vigilancia. Está basado en el sensor propuesto por Lichsteiner y sus colegas [10] y mejora algunas de
sus características como el ancho de banda, el tiempo de latencia y el fill factor. Para ello, hace
uso de amplificadores tras el fotoreceptor. De esta forma, se aumenta la ganancia y se reduce el
tamaño del píxel, ya que la ganancia del circuito derivador no necesita ser tan alta, y, por tanto, las
capacidades del mismo pueden hacerse más pequeñas. En el circuito propuesto por Lichsteiner la
ganancia del píxel estaba basada únicamente en el cociente entre capacidades del circuito diferenciador. Por esta razón, grandes capacidades eran requeridas para obtener una sensibilidad al contraste aceptable.
En cuanto al trabajo futuro, después de haber testado exhaustivamente los chips fabricados,
se pueden proponer varias mejoras. En primer lugar, el calibrado de la retina podría hacerse en el
propio chip de forma automática. Hasta ahora el calibrado se había hecho off-line según el procedimiento descrito en [13]. Sería perfectamente factible, implementar un controlador en el propio
chip que lo realice de forma automático. Simplemente, sería necesario iluminar el chip de forma
uniforme e implementar un algoritmo que escoja las palabras de calibrado que minimizan las frecuencias de salida de cada píxel.
En cuanto a la retina de contraste espacial, el fill factor obtenido (2%) es relativamente bajo.
Podría mejorarse si se reduce el número de bits de las palabras de calibrado (3 ó 4 bits sería suficiente) porque el valor del FPN obtenido (0.6%) es muy bajo. Ello posibilitaría implementar una
retina de contraste espacial de mayor resolución.
Finalmente, la retina de contraste temporal hace uso de amplificadores en inversión fuerte.
Ello hace que el consumo sea ligeramente superior al del diseño de Lichsteiner [10]. En el futuro,
sería posible implementar un diseño similar con transistores operando en inversión débil para
reducir el consumo del chip.
44
Sistemas de Visión AER
45
Sistemas de Visión AER
46
Bibliografía
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
M. Sivilotti, “Wiring considerations in analog VLSI systems with application to field-programmable networks,” Ph.D. dissertation, Calif. Inst.Technol., Pasadena, 1991.
M. Mahowald, “VLSI analogs of neural visual processing: a synthesis of form and function” Ph.D. dissertation, California Institute of Technology, Pasadena, 1991.
T. Delbrück, Library essentials, Analog VLSI and neural systems by Carver Mead, Addison
Wesley, 1986. The Neuromorphic Engineer, 1(1):11, 2004. http://ine-web.org/research/
newsletters/index.html.
Kwabena A. Boahen, “Retinomorphic Vision Systems: Reverse Engineering the Vertebrate
Retina”, Ph. D. dissertation, California Institute of Technology, Pasadena, 1996.
G. M. Shepherd, The Synaptic Organization of the Brain, 3rd ed. Oxford, U.K.: Oxford University Press, 1990.
M. Mahowald. An Analog VLSI System for Stereoscopic Vision, Kluwer, Boston, MA, 1994.
U. Mallik, et al., “Temporal Change Threshold Detection Imager”, in ISDD Dig. of Tech.
Papers, San Francisco, 2005, pp362-363.
Y. Chi et al., “CMOS Camera with In-pixel Temporal Change Detection and ADC”, IEEE
Journal of Solid State Circuits, vol. 42, pp, 2187-2196, OCT 2007.
Patrick Lichsteiner et all, “A 128x128 120dB 15 µs Latency Asynchronous Temporal Contrast Vision Sensor”, IEEE Journal of Solid State Circuits, vol. 43, pp. 566-576, 2008.
P. Lichsteiner, et al., “A 128x128 120dB 30mW Asynchronous Vision Sensor that Responds
to Relative Intensity Change”, in ISSCC Dig. of Tech. Papers, San Francisco, 2006, pp.
508-509 (27.9)
S. Thorpe, D. Fize, C. Marlot, “Speed of Processing in the Human Visual System”, Nature
381:520-2, 1996.
X. G. Qi, X.; Harris J., “A Time-to-first-spike CMOS imager”, in Proc. of the 2004 IEEE
International Symposium on Circuits and Systems (ISCAS04), Vancouver, Canada, 2004,
pp. 824-827.
J. A. Leñero-Bardallo, T. Serrano-Gotarredona, and B. Linares-Barranco, “A Calibration
Technique for Very Low Current and Compact Tunable Neuromorphic Cells. Application to
47
BIBLIOGRAFÍA
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
5-bit 20nA DACs”, IEEE Transactions on Circuits and Systems, Part-II: Brief Papers,
vol.55, No. 6, pp. 522-526, June 2008.
J. Costas-Santos, T. Serrano-Gotarredona, R. Serrano-Gotarredona and B. Linares-Barranco, “A Spatial Contrast Retina with On-chip Calibration for Neuromorphic Spike-Based
AER Vision Systems”, IEEE Trans. Circuits and Systems, Part-I: Regular Papers, vol. 54,
No. 7, pp. 1444-1458, July 2007.
M. Barbaro, P.-Y. Burgi, A. Mortara, P. Nussbaum, and F. Heitger. A 100100 Pixel Silicon
Retina for Gradient Extraction with Steering Filter Capabilities and Temporal Output Coding. IEEE Journal of Solid-State Circuits, 37(2):160-172, Feb. 2002.
P. F. Ruedi, et al., “A 128x128 Pixel 120-dB Dynamic-range Vision-Sensor Chip for Image
Contrast and Orientation Extraction”, IEEE Journal of Solid-State Circuits, 38:2325-2333,
Dec. 2003.
P. F. Ruedi, et al., “An SoC Combining a 132dB QVGA Pixel Array and a 32b DSP/MCU
Processor for Vision Applications”, in IEEE ISSCC Dig. of Tech. Papers, 2009, pp.46-47,
47a.
Honghao Ji and Pamela A. Abshire, “Fundamentals of Silicon-Based Phototransduction” in
“CMOS Imagers from Phototransduction to Image Processing”, Kluwer Academic Publishers.
E. Culurciello, R. Etienne-Cummings, and K. A. Boahen, “A Biomorphic Digital Image
Sensor”, IEEE J. Solid-State Circuits, vol.38, pp.281-294, 2003.
S. Chen, and A. Bermak, “Arbitrated Time-to-First Spike CMOS Image Sensor with OnChip Histogram Equalization”, IEEE Transactions VLSI Systems, vol. 15, No. 3, pp 346357, March 2007.
K. A. Zaghloul and K. Boahen, “Optic Nerve Signals in a Neuromorphic Chip: Part 1”,
IEEE Transactions on Biomedical Engineering, vol 51, pp. 657-666, 2004.
K. A. Zaghloul and K. Boahen, “Optic Nerve Signals in a Neuromorphic Chip: Part 2”,
IEEE Transactions on Biomedical Engineering, vol 51, pp. 667-675, 2004.
J. A. Leñero-Bardallo, T. Serrano-Gotarredona, and B. Linares-Barranco, “A 5-Decade
Dynamic Range Ambient-Light-Independent Calibrated Signed-Spatial-Contrast AER Retina with 0.1ms Latency and Optional Time-to-First-Spike Mode”, Transactions on Circuits
and Systems, Part I, Under revision.
C. Posch, et al., “High DR, Low Date-rate Imaging Based on an Asynchronous, Self-triggered Address-event PWM Array with Pixel-level Temporal Redundancy Suppression”, in
2010, in ISCAS 2010.
R. R. Harrison, J.A. Bragg, P. Hasler, B.A. Minch, and S.P. DeWeerth, “A CMOS
programmable analog memory-cell array using floating-gate circuits,” IEEE Trans. on
48
BIBLIOGRAFÍA
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
Circuits and Systems, Part II, vol. 48, No. 1, pp. 4-11, Jan. 2001.
Y. L. Wong, M. H. Cohen, and P. A. Abshire, “128x128 floating gate imager with selfadapting fixed pattern noise reduction,” Proc. of the IEEE 2005 Int. Symp. on Circuits and
Systems (ISCAS’05), vol. 5, pp. 5314-5317, 2005.
R. Serrano-Gotarredona, T. Serrano-Gotarredona, A. Acosta-Jimenez, and B. LinaresBarranco, “A Neuromorphic Cortical-Layer Microchip for Spike-Based Event Processing
Vision Systems,” IEEE Trans. on Circuits and Systems I: Regular Papers, vol. 53, No. 12,
pp. 2548-2566, Dec. 2006.
B. Linares-Barranco, T. Serrano-Gotarredona, and R. Serrano-Gotarredona, "Compact LowPower Calibration Mini-DACs for Neural Massive Arrays with Programmable Weights,"
IEEE Trans. on Neural Networks, vol. 14, No. 5, pp. 1207-1216, September 2003.
C. A. Laber, C. F. Rahim, S. F. Dreyer, G. T. Uehara, P. T. Kwok, and P. R. Gray, “Design
Considerations for a high-Performance 3mm CMOS Analog Standard-Cell Library,” IEEE
J. Solid-State Circuits, vol. SC-22, No. 2, pp. 181-189, April 1987.
T. Delbrück and A. Van Shaik, “Bias Current Generators with Wide Dynamic Range,” Int.
Journal of Analog Integrated Circuits and Signal Processing, No. 43, pp. 247-268, June
2005.
R. R. Harrison, J.A. Bragg, P. Hasler, B.A. Minch, and S.P. DeWeerth, “A CMOS
programmable analog memory-cell array using floating-gate circuits,” IEEE Trans. on
Circuits and Systems, Part II, vol. 48, No. 1, pp. 4-11, Jan. 2001.
K. Bult and G.J.G.M. Geelen, “An inherently linear and compact MOST-only current
division technique,” IEEE J. Solid-State Circuits, vol. 27, No. 12, pp. 1730-1735, Dec.
1992.
T. Serrano-Gotarredona and B. Linares-Barranco, "CMOS Mismatch Model valid from
Weak to Strong Inversion", Proc. of the 2003 European Solid State Circuits Conference,
(ESSCIRC’03), pp. 627-630, September 2003.
P.R. Gray, P.J. Hurst, S.H. Lewis, and R. G. Meyer, Analysis and Design of Analog
Integrated Circuits, 4th Edition, John Wiley, 2001.
Rafael Serrano-Gotarredona, Luis Camuñas-Mesa, Teresa Serrano-Gotarredona, Juan A.
Leñero-Bardallo and Bernabé Linares-Barranco, “The Stochastic I-Pot: A Circuit Block for
Programming Bias Currents”, IEEE Transaction on Circuits and Systems-II: Brief Papers,
vol 19, No. 7, pp. 1196-1219, July 2008.
B. Linares-Barranco and T. Serrano-Gotarredona, “On the Design and Characterization of
Femtoampere Current-Mode Circuits”, IEEE Journal of Solid-State Circuits, vol. 38, No. 8,
pp. 1353-1363, August 2003.
M. Azadmehr, J. Abrahamsen, and P. Häfliger, “A Foveated AER Imager Chip”, in Proc. of
49
BIBLIOGRAFÍA
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
the IEEE International Symposium on Circuits and Systems (ISCAS2005), pp. 2751-2754,
Kobe, Japan, 2005.
R.J. Vogelstein, U. Mallik, E. Culurciello, R. Etienne-Cummings, and G. Cauwenberghs,
“Spatial acuity modulation of an address-event imager,” Proc. of the 2004 11th IEEE
International Conference on Electronics, Circuits and Systems (ICECS 2004), pp. 207-210,
Dec. 2004.
J. Kramer, “An Integrated Optical Transient Sensor”, IEEE Transactions on Circuits and
Systems, Part II: Analog and Digital Signal Processing, v. 49, No. 9, pp. 612-628, Sep.
2002.
M. Arias-Estrada, D. Poussart, and M. Tremblay, “Motion Vision Sensor Architecture with
Asynchronous Self-Signaling Pixels”, Proceedings of the 7th International Workshop on
Computer Architecture for Machine Perception (CAMP07), pp. 75-83, 1997.
K. Boahen, “Retinomorphic Chips that see Quadruple Images”, Proceedings of International Conference of Microelectronics for Neural, Fuzzy and Bio-inspired Systems
(Microneuro99), pp. 12-20, Granada, Spain, 1999.
S. Thorpe, et al., “SpikeNet: Real-time Visual Processing in the Human Visual System”,
Neurocomputing 58-60: 857-64, 2004.
K. Boahen and A. Andreou, “A Contrast-Sensitive Retina with Reciprocal Synapses”, in J.
E. Moddy (Ed.), Advances in neural information processing, vol. 4, pp- 764-772, San
Mateo CA, 1992. Morgan Kaufman.
A. G. Andreou and K. Boahen, “Translinear Circuits in Subthreshold CMOS”, Analog Integrated Circuits and Signal Processing, Kluwer, no. 9 , pp. 141-166, Apr. 1996.
K. Boahen, “Point-to-Point Connectivity Between Neuromorphic Chips Using Address
Events”, IEEE Transactions on Circuits and Systems Part-II, vol. 47, No. 5, pp. 416-434,
May 2000.
R. Berner, T. Delbruck, A. Civit-Balcells and A. Linares-Barranco,“A 5 Meps $100 USB2.0
Address-Event Monitor-Sequencer Interface”, IEEE International Symposium on Circuits
and Systems,2007, ISCAS 2007.
jAER Open Source Project 2007 [Online]. Available: http://jaer.wiki.sourceforge.net.
V. Gruev and Etienne-Cummings, R. A. pipelined temporal difference imager. IEEE J. Solid
State Circuits, 39:538-543, 2004.
M. A. Mahowald, “VLSI Analogs of Neuronal Visual Processing: A Synthesis of Form and
Function”, PhD, Computation adn Neural Systems, Caltech, Pasadena, California, 1992.
D. A. Baylor, T. D. Lamb, and K.W. Yau, “Responses of Retinal Rods to Single Photons”, J.
Physiology., 288:613-634, 1979.
D. A. Baylor, B. J. Nunn, and J. L. Schnapf. “The Photo Current, Noise, and Spectral Sen-
50
BIBLIOGRAFÍA
sitivity of Rods of the Monkey Macaca Fascicularis”. J. Physiology, 357:575-607, 1984.
51
ANEXOS
53
University of Sevilla
Department of Electronics and Electromagnetism
Ph.D. Dissertation
Study, Design, Implementation, and Test of VLSI
Retinae Sensitive to Spatial and Temporal Contrast
By Juan Antonio Leñero Bardallo
Advisors:
Dr. Bernabé Linares Barranco
Dr. Teresa Serrano Gotarredona
Seville, 2010
56
To my parents
57
58
ACKNOWLEDGEMENTS
This dissertation is the result of my work at the Sevilla Microelectronics
Institute during four years. I would like to be thankful to my thesis advisors,
Bernabé Linares and Teresa Serrano. They have helped me and they have conducted my work during all the time I have been elaborating my thesis. Without
their help, this document would not have been possible.
All my work at the Sevilla Microelectronics Institute during this time has
been funded by a JAE scholarship. Thanks to this program, I also had the opportunity of spending two months in the University of Oslo, researching at the
Department of Informatics with Dr. Philipp Häfliger.
I also would like to mention my parents and brothers for their support and
care during all this time.
This work was made with the collaboration of several people. I have to
mention the ATC group of the University of Sevilla for their useful boards and
software that I used to test the sensors, Philipp Häfliger for the supplied Caviar
board and the lens mount, Tobi Delbrück for the jAER software and the supplied
temporal contrast sensor, and the anonymous reviewers that have contributed to
improved the quality of this document and our publications.
Finally, I have to thank my colleagues at the Sevilla Microelectronics Institute. I have learnt a lot of valuable skills from them. With their company, all this
time in Sevilla has been really enjoyable.
59
60
OUTLINE
Nowadays, vision systems are commonplace and widely used to perform plenty of different
tasks. For example, we are recorded when we park our car in a car place or when we walk though
an airport. Everybody is used to such kind of devices. There is also an huge potential market, specially related to cell phones. Almost all of them have cameras nowadays. All these imagers are
based on frames (frame-based devices). With regard to artificial vision, we usually speak about
machine vision systems (systems made up by cameras and actuators) that also perform specific
tasks like surveillance or remote control. However, such devices work in very controlled environments in a repetitive way.
Frame-based devices have inherent limitations. Their bandwidth is limited by the sample
rate. They are always transmitting data periodically that can be very redundant if there is no
changes in the scene (frames are always transmitted if the system is working). They do not process the information and a lot of redundant data is involved. Their dynamic range is reduced
because AGC (Automatic Gain Control) is global to the pixel matrix. Each pixel has the same
gain and the same integration time. The capacity to integrate charge is also limited. In unsupervised environments, bandwidth and dynamic range can limit the performance of the whole vision
system.
On the other hand, biological systems are frameless. They are massively parallel and data
driven. Such systems are structured in layers. Each layer performs one specific task and sends the
relevant information to the next layers. Eyes are optimized to extract the relevant information
from the visual scene. If we compare conventional systems to bio-inspired ones, the advantages of
the second ones are obvious.
Neuromorphic engineering is based on the belief that emulating such biological systems, we
could develop electronic devices with similar features. The rise of VLSI systems leads to systems
with pixels with an acceptable fill factor and level of complexity.
In this thesis, we will describe two AER (Address Event Representation) vision systems that
imitate two specific retinal functionalities. The first one detects the spatial contrast. The second
one is sensitive to temporal contrast. Both of them are spike-based devices and they have two separate channels to process the negative and positive contrast (signed output). Detection of spatial
and temporal contrast are two very important tasks that the human retina can perform. With the
information of the spatial contrast, we can detect edges and shapes. With the detection of the temporal contrast, we can detect movement, relative depths and group objects.
The first sensor (described in chapter 4) overcomes some of the drawbacks of prior AER
spatial contrast imagers. It uses a new calibration system to reduce mismatch, it also has signed
61
OUTLINE
output, and a new thresholding mechanism that allows removal of residual mismatch after calibration. It also has an optional TFS (time-to-first spike) mode that combines the advantages of framebased systems and AER ones. Its main features are very low FPN (0.6% of the maximum possible
output frequency when the sensor detects spatial contrast), good response time (0.1-10ms), high
sensitivity and very low power consumption (0.66-6.6mW). A test prototype of 32 x 32 pixels was
fabricated.
The second sensor (described in chapter 5) is specially suitable for high speed applications.
It has very low latency (down to 3.5µs ) and high pixel bandwidth (up to 8KHz). It uses amplifiers after the logarithmic photoreceptor to reduce the retina response time and increase its sensitivity to temporal contrast. This prototype has a fill factor of 8% and has 128 x 128 pixels. It also has
two independent channels to process positive and negative events independently.
The thesis is structured into 6 chapters. Chapter 1 describes current trends in imager design
and compares the features of conventional imagers to the ones of bio-inspired sensors. It also
explains the state of the art and establishes a comparison between the more relevant AER imagers
presented during the last few years. Chapter 2 presents the Stochastic I-Pot, a powerful circuit to
generate precisely programmable bias currents. Chapter 3 explains thoroughly the new calibration
system and its possible applications to reduce mismatch effects in large neuromorphic arrays.
Chapter 4 presents the new AER spatial contrast vision sensor. The complete circuitry is
explained. Detailed experimental results are also provided. Chapter 5 describes the new AER
temporal contrast vision sensor. Circuitry and experimental results are displayed. Finally, in chapter 6, we draw some conclusions about the present and the possible future work.
62
CHAPTER 1
Introduction
1.1 Conventional Vision Systems Based on Frames
In the last ten years, sales of digital cameras have grown incredibly. The huge potential market and the raise of cell phones has driven a very fast development in electronic imager design.
This has also led to improved and cheaper imagers, with very small pixels. This tendency is continuing. Nowadays, we usually speak about the number of megapixels as a figure of merit.
All these conventional vision systems are based on frames [52]. A frame is an 2D matrix
with elements that represent the intensities and color of the visual scene under observation during
a period of time. Video data consists of a stream of frames taken with a fixed frequency. Framebased systems are commonplace and sometimes when we speak about vision systems they are
taken for granted. Today’s CCDs and CMOS imagers are frame-based devices.
The main advantage of these devices is that they use simple and small size pixels. A small
pixel implies high resolution, large fill factor and low cost. Such advantages are inherent to framebased devices. Another point to consider is that frames are compatible with all the commercial
frame-based displays.
However, frames have some drawbacks to take into account. Pixel bandwidth is limited by
the frame rate. In conventional systems, there is a trade-off among system bandwidth, frame rate,
sensor sensitivity, and the scan period. The frame rate ( 1 ⁄ T frame ) is the number of frames transmitted per unit of time. System bandwidth is the maximum amount of data that our system can
process per unit of time. The exposition time, Texpositon, is the amount of time that pixels are collecting charge to generate one frame ( T exp osition ≤ T frame ) and we usually can adjust it. Sensitivity is the minimum detectable illumination level during the scan period. Frame-based systems
accumulate charge (in a capacitor or a photogate) during the scan period. After this, charge at the
photogate or the capacitor is read and resetted. If the scan period is too low, the amount of charge
integrated will be very low. For this reason, typically conventional systems usually need to work
in environments with high and controled illumination conditions to achieve high speed without
loosing sensitivity. If we increase frame rate, the amount of output data will be increased. We cannot increase the frame rate arbitrarily because system bandwidth limits the amount of data that the
63
Chapter 1: Introduction
N-1
N
Fig 1.1: Structure of a CCD device with N photogates. We can notice how the charge
packet is transferred from one side to the another one.
sensor can process. Temporal changes in the scene cannot be detected in the time between two
different frames.
Another disadvantage is redundancy. Pixels are continuously scanning and sending information, even if there is no new information to transmit. This leads to a huge load of redundant data.
Finally, AGC (Automatic Gain Control) is usually global in frame-based systems. Thus, gain control is not properly adjusted when there is intrascene variations of illumination. For this reason, it
is very important to create a good scene illumination in machine vision. In conclusion, limited
bandwidth, identical gain for all the pixels and redundancy are assumed for conventional vision
systems.
1.1.1 Charge-Coupled Devices
During several years the predominant technology for optical imaging applications has been
the charge-coupled device (CCD) [18]. CCD technology revolutionized the field of digital imaging by enabling diverse applications in consumer electronics, scientific imaging and computer
vision. Its advantages are high sensitivity, high resolution, large dynamic range, and large array
size. CCD is a very mature and developed technology. CCD imagers provide images with excellent quality and are still the dominant imaging technology.
A Charge-Coupled Device (CCD) is an image sensor (see Figure 1.1), consisting of an integrated circuit containing an array of linked, or coupled, MOS capacitors sensitive to light. MOS
capacitors (also called photogates) are operated in deep depletion on a continuous insulator layer
over the semiconductor substrate. Nowadays, pinned photodiodes are being used instead photogates. Under the control of an external circuit, each capacitor can transfer its electric charge to one
of its neighbors as can be seen in Figure 1.1.
There are three conventional architectures: full-frame, frame-transfer and interline. They
basically differ on their approach to the problem of shuttering or how the charge is transferred
from the photogates.
In a full-frame device, all of the image area is active and there is no electronic shutter. A
mechanical shutter must be added to this type of sensor or the image will smear (accumulated
charge loss) as the device is clocked or read out. It is a useful choice when light has to be collected
accurately.
64
Chapter 1: Introduction
Frame-transfer architectures are made with an opaque mask (typically aluminium). Half of
the pixel area is covered with this mask. The charge is quickly transferred from the photogates to
this opaque region with little smear percent. Once the charge has been transferred, it can be read
out slowly and the active area can be integrating a new frame at the same time. The main drawback of this architecture is that it requires twice the silicon area than a full-frame device.
In the interline architecture there is a mask over each column of the image sensor, so only
one pixel shift has to occur to transfer charge from the image area to the storage area. The disadvantage of the scheme is that fill factor is reduced 50%. Adding special microlenses over the
opaque mask, light can be directed away of the opaque region to the active area increasing the fill
factor. With this approach, charge can be transferred in the order of microseconds with a very low
smear. This is the dominant architecture that we can find in consumer CCD cameras.
The choice of architecture comes down to one of utility. If the application cannot tolerate an
expensive, failure-prone, power-intensive mechanical shutter, an interline device is the right
choice. Consumer snap-shot cameras have used interline devices. On the other hand, for those
applications that require the best possible light collection and issues of money, power and time are
less important, the full-frame device is the right choice. Astronomers tend to prefer full-frame
devices. The frame-transfer falls in between and was a common choice before the fill-factor issue
of interline devices was addressed. Today, frame-transfer is usually chosen when an interline
architecture is not available, such as in a back-illuminated device.
1.1.2 The CMOS Sensor
CMOS sensors, also known as Active Pixel Sensors (APS), have been studied since the
1980s as an alternative to CCD technology. APS technology is now beginning to enable new
applications in digital imaging by offering improved performance relative to CCD technology in
the areas of low cost, low-power operation, high speed and ease of integration.
An APS is an imager in which every pixel includes at least one active transistor. Transistors
may operate as both amplifier and buffer to isolate photogenerated charge from the large capacitance of the common output line. Any photodetector can be used for this approach: photo diodes
or photogates [18]. APS pixels usually operate in charge integration mode, but may operate with
photo transistors as well. Pixels are implemented with circuitry shared by pixels in the same row,
and other circuitry by the same column. At the end of each column there is an integrated amplifier
to read out the selected pixel.
Figure 1.2 shows an APS with a photo diode. Light causes an accumulation of charge on the
capacitance of the photo diode photo current that depends on light intensity. Transistor, M1, is a
switch that resets the device. M2 is a source follower that allows to observe the pixel voltage without removing the accumulated charge. M3 acts as a switch to select one single pixel of one column
to be read.
The primary motivation for developing CMOS-based imagers is the ease of integration and
the ability to incorporate signal conditioning circuitry and even image processing circuitry into
the same device. APS systems offer compact size, light weight, low cost, and low power consumption. Despite these advantages, CCD still remains the dominant imaging technology [18].
The main reason is that CCD is a very mature technology with specialized and optimized fabrication processes that provide images with excellent quality, higher sensitivity and higher dynamic
range.
65
Reset
M1
M2
M3
Vbias
Column Bias
Chapter 1: Introduction
Row selection
M4
Fig 1.2: APS photocircuit using a photo diode. The components inside the dashed-line
box constitute the pixel [18].
1.1.3 Color Cameras
Digital color cameras generally use a Bayer mask over the CCD or the APS as is depicted in
Figure 1.3. Each square of four pixels has one filter red, one blue, and two green (the human eye
is more sensitive to green than either red or blue). The result of this is that luminance information
is collected at every pixel, but the color resolution is lower than the luminance resolution.
Better color separation can be reached by three-CCD devices (3CCD) and a dichroic beam
splitter prism, that splits the image into red, green and blue components. Each of the three CCDs
is arranged to respond to a particular color. Some semi-professional digital video camcorders (and
most professional camcorders) use this technique. Another advantage of 3CCD over a Bayer
mask device is higher quantum efficiency (and therefore higher light sensitivity for a given aperture size). This is because in a 3CCD device most of the light entering the aperture is captured by
a sensor, while a Bayer mask absorbs a high proportion (about 2/3) of the light falling on each
CCD pixel.
The Foveon X3 sensor [55] presents another approach to obtain color images. It is a CMOS
image sensor for digital cameras, designed by Foveon, Inc. It uses an array of photosites, each of
which consists of three vertically stacked photodiodes, that are organized in a two-dimensional
grid (as is shown in Figure 1.4). Each of the three stacked photodiodes responds to different wavelengths of light, because they have a different spectral sensitivity curve. This is due to that fact
that different wavelengths of light penetrate silicon to different depths. The signals from the three
photodiodes are then processed, resulting in data that provides the three additive primary colors,
red, green, and blue.
66
Chapter 1: Introduction
(a)
(b)
Fig 1.3: (a) The Bayer arrangement of color filters on the pixel array of an image sensor. (b)
Cross-section of the sensor.
Fig 1.4: Color absorption in silicon and the Foveon X3 sensor.
67
Chapter 1: Introduction
1.2 Bio-inspired Systems and Neuromorphic Engineering
If we compare biological vision systems to frame-based systems, the first ones have clear
advantages over the artificial ones. Biological vision systems are very complex and have amazing
capabilities [4]. The visual system contains about 100 million photoreceptors. It has a dynamic
range of more than 10 decades [4] that allows to detect dim starlight (it can sense the absorption
of one single photon per integration time) to direct sunlight (it saturates at a 100 photons per integration time) [52]-[53]. Under ideal conditions, humans can detect a 0.5-percent change in mean
intensity [56]. The human retina can resolve 1 arc minute (1/60 degree) [57] at the fovea, a small
specialized region in the center of the retina where the cones are extremely small and are packed
densely. Our eyes are also sensitive to temporal changes in the image. The integration time of the
primate visual system is on the orther of 0.1 second [58]. Our eyes are sensitive to the wavelength
of light. The human eye has several 100 million photoreceptors and the density of photoreceptors
in the center of the visual fields as high as 150.000 photoreceptors per square-millimeter [4]. Photoreceptor pitch goes down to about 2.6µm .
Looking at the state of the art, machine vision has only certain features that are better when
they are used in very controlled environments. Biological vision systems perform efficiently real
time tasks under very different conditions and are quite different to frame-based systems. Biological systems are not based on frames. They are continually detecting information. They are also
massively parallel and data driven. Such systems present low redundancy, high speed response
and high dynamic range. Human eyes has also high power efficiency. However, their most important feature is the possibility of extraction of relevant information from the visual scenes and
transmit it to the brain in an accurate way. They are estructurated in layers of neurons connected
between them. Each layer perform a certain task like spatial or temporal contrast detection. To
extract the relevant information of the scene, a spatio-temporal filtering operation is done. This
way, temporal and spatial redundancy is removed. Figure 1.5 shows an example of the operation
of spatial filtering done in the human eye before transmitting the visual data to the brain. On the
left side, there is on image captured with a conventional system. On the left side, there is the same
image after filtering it with a high pass filter. We can notice that edges and shapes have been
enhanced. At the same time, all the redundant data that was not necessary to distinguish shapes
and edges was removed. Developing bio-inspired systems that performs any kind of processing
the bandwidth usage will be optimized.
Figure 1.6 illustrates how frame-based and bio-inspired ones performs when they detect a
very fast rotating stimulus and some advantages of bio-inspired systems. We have plotted with
points the digital recorded output of one of our temporal contrast retinas when it was stimulated
with a high speed rotating dot. Only pixels that detected temporal contrast generated outputs. In
biological systems there are not scan periods and pixels are continuously detecting changes in the
scene. If pixel bandwidth is enough high, the rotating dot will be detected. As we mentioned previously, biological systems usually process the information. In this particular case, the sensor was
able to detect the temporal contrast (temporal filtering) and its sign (each color represents one
sign of the spatial contrast). We have also plotted how a frame-based system would perform. It
would be generating frames periodically. Such frames would sense the accumulated photo charge
of all the pixels of the retina without processing the information. In this particular case, Tframe is
higher than the stimulus period. This would be a serious limitation and we would not detect the
68
Chapter 1: Introduction
(a)
(b)
Fig 1.5: Example of the importance of spatial filtering to detect shapes and edges and
remove redundancy. (a) Original image captured with a frame-based system. (b) Resulting image after high pass spatial filtering by using Matlab.
69
Chapter 1: Introduction
t
y
x
10.000 rps
Bio-inspired
system outputs
Rotating dot
stimulus
Tframe
Frame-based system outputs
Fig 1.6: Comparison between a conventional system based on frames and a bio-inspired one
tested in our laboratories. The first one (blue points) does not process the information and
cannot detect the stimulus because its period is lower than Tframe. All the matrix pixels
transmit the sensed luminance level. The second one (yellow and cyan points) extracts the
temporal contrast (with sign) and can detect the rotating dot. In this case, only regions with
contrast transmit information. A temporal filtering operation was done by the sensor.
70
Chapter 1: Introduction
(a)
(b)
Fig 1.7: (a) Image with very different levels of illumination taken with a conventional camera with a global AGC. (b) Same image after processing it with an algorithm that implements a local AGC.
dot movement. Each frame would detect a circle, but not the dot positions. All the frames would
be identical.
Another advantage of bio-inspired systems is the local AGC (Automatic Gain Control). The
operation range of a typical imager is about 10 decades. Pixel gain has to be adjust to the illumination level to have always the same sensitivity and response in the same way to the relative illumination variations. Conventional systems have a global AGC mechanism. The average chip
luminance level is sensed and all the pixel gains are adjusted in the same way. However, this
mechanism does not offer good results when there are regions in the pixels matrix with very different levels of illumination. On the contrary, biological systems have a local AGC. Each individual pixel senses the local photo current and its sensitivity is adjusted according this sensed
illumination value. Thus, it is possible to capture images where there are very different levels of
illumination without loosing sensitivity in any region of the scene. Figure 1.7(a) shows a image
taken with a conventional camera with a global AGC. In the image, there were regions with very
different levels of illumination. We can notice that it is not possible to distinguish the objects and
shapes beyond the window. Figure 1.7(b) shows the some picture after using an algorithm that
sets each pixel luminance value taking into account only its neighboring pixels (local AGC) and
not all the matrix pixels. The algorithm increases the level of contrast of the scene. In this case, it
is possible to distinguish the objects beyond the window.
For the previously mentioned reasons, it makes sense to try to develop devices that emulate
the biological systems. Neuromorphic engineering is a discipline that seeks to implement neural
structures in silicon. Primate brains are structured in layers of neurons. In each layer, neurons are
connected to a large number of neurons (~104) from the next layer [5]. There are massive connections that allows to implement efficient processing of the information between layers. In early
neuromorphic engineering, it soon became clear that one of the major differences between brains
71
Chapter 1: Introduction
Fig 1.8: AER interchip point to point communication scheme.
and computers was the connectivity between processing units. In order to try to develop multichip
systems connected between them massively in a accurate way, efficient mechanisms of interconnection and specific communication protocols are required. Nowadays, VLSI circuit technologies
allows to fabricate on a single chip thousands of neurons. However, it is not viable and possible to
connect all of them physically. It has also sense to implement layers in different chips leading to
multichip systems. In such systems, each single chip (layer) run in parallel and the single components are more simple.
AER interchip communication was originally proposed by Mahowald and Silvilotti [1]-[2].
Biological neurons of a layer communicate with neurons of the next layer by sending information
with spikes. In Fig 1.8, we can see a scheme of an AER communication between two different
chips that implement two different layers of neurons. One chip initiates the communication sending a Request Signal (RQST) to the other chip. If the receiver is ready to receive information, it
sends an Acknowledge (ACK) signal to the first chip. Then the first chip sends the information
with the address of the neuron that has to receive the spike. This way, different neurons in different chips can communicate without being physically connected. Arbitration mechanisms are necessary in the transmitter to avoid collisions between neurons in the fast digital bus.
Multi-chip systems are usually made up with flexible systems that can be built independently to meet specific needs. They also allow to combine the effort of different groups avoiding
the chip interconnection problems. AER can be considered a real time communication system
because the total activity of a population of sensory or computational units is not close to the
bandwidth of the communication channel.
1.3 The Biological Retina
Fig 1.9 shows a cross-section of the human eye and schematic of the different cells that compound the retina. Light enters the eye passing through the cornea. Around the lens there is a colored ring of muscle called the iris. The iris blocks light, forming an adjustable aperture in the
center where light can enter the eye through the lens. This hole, called the pupil, is adjusted by the
iris muscles to maintain a relatively constant level of light entering the eye. The light passing
72
Chapter 1: Introduction
Sclera
Retina
Bipolar cell
Rod
Ganglion cell
Cone
Cornea
Fovea
Lens
Pupil
Iris
Amacrine cell
Optic nerve
Horizontal cell
Optic disc
Fig 1.9: Human eye and schematic of the retina.
through the pupil is focused by the cornea and the lens. While the cornea has a fixed curvature,
the lens can be stretched to adjust the focus to the distance of the object which the eye is looking
at. The innermost tissue layer on the back portion of the eyeball is called the retina. All the individual components through which light travels within the eye before reaching the retina are transparent, minimizing attenuation of the light. As we can see in Fig 1.9, the retina contains different
kind of cells: rods, cones, and neurons. Its main parts are the fovea, the optical disc and the optical
nerve. The fovea is a small specialized region in the center of the retina where the cones are
extremely small and packed densely (there are no rods in this region). The optical disc is a point
on the retina where the axons of the ganglion cells leave the eye to form the optic nerve. There are
no photosensitive cells at the optic disc.
The retina is structured in two layers: the outer plexiform layer (OPL) and the inner plexiform layer (IPL). The OPL is the more pheriferical layer and the IPL is the more central one.
The retina contain several kinds of cells and photoreceptors. There are two different kinds of
photoreceptors [4], called rods and cones. They operate under different conditions of illumination.
Rods operate in dim light and they cannot distinguish colors. Cones are less sensitive and operate
under bright light and are sensitive to the wavelength of light due to the presence of three different
types of cone pigments, with peak absorbances at 420nm (violet), 530nm (green-yellow) and
560nm (blue-green). In comparison, rods are tuned to 500nm (blue-green).
Neural responses go from the cones and rods to a series of linking cells, called bipolar, horizontal, amacrine cells, and ganglion cells. Bipolar cells are neurons placed between the OPL and
the IPL and have a bipolar structure. Horizontal cells play an inhibitory role. Primates have two
horizontal-cell types with no color selectivity. Amacrine cells are also inhibitory. Ganglion cells
are the only output channel of the retina. These cells communicate by sending trains of impulses.
All these cells are connected creating a neural network that provide a mechanism for spatial
and temporal filtering that allows to emphasize edges and temporal changes in the visual field.
73
This network enhances the relevant information and does not send to the brain redundant or useless information.
We can conclude that biological vision is not just about measuring or sensing the power and
the light spectral composition of the visual scene. Biological systems can also process the information and enhance the useful information to send to the brain and avoid redundancy.
1.4 AER Vision Sensors
1.4.1 Adaptive Temporal and Spatial Filtering
The retina has evolved sophisticated filtering and adaptation mechanisms to reduce redundancy and to improve coding efficiency [4]. Some of these mechanisms are local automatic gain
control, bandpass spatio-temporal spatio-temporal filtering and highpass temporal and spatial filtering. Bandpass spatio-temporal filtering in the first stage of the retina (OPL) attenuates signals
that do not occur at a fine spatial or temporal scale, removing redundant transmission of low-frequency signals and eliminating noisy high frequency signals. Highpass temporal and spatial filtering in the second stage of the retina (IPL) attenuates signals that do not occur at a fine spatial scale
and temporal scale, eliminating the redundant signals passed by the OPL, which responds
strongly to low temporal frequencies that occur at high spatial frequencies (sustained response to
static edge) or to low spatial frequencies that occur at high temporal frequencies (blurring of rapidly moving edge).
Boahen developed a simple dynamic model of the retina [4] to study the spatio-temporal
behavior. It was a physical model made up resistors, capacitors, and transconductances. It was
based on the neurocircuitry of the vertebrate retina and included several major synaptic interactions in the outer plexiform layer (OPL). This model was very useful to understand and study the
spatio-temporal filtering in the retina, but it did not have useful applications because it has so
many practical difficulties.
1.4.2 Spatial Contrast Extraction
There have been several attemps to implement a spatial contrast sensitive retina. The first
functional retina able to detect contrast [46] was proposed by K. Boahen and A. Andreou. That
retina could measure the spatio-temporal contrast. It was based on the model [4] previously proposed by Boahen to study the spatio-temporal filtering in the human retina. The spatio-temporal
retina had a resolution of 90 x 92 pixels and a dynamic range of 5 decades. It mains drawbacks
were mismatch and bandwidth consumption when there was no contrast. It could not detect the
sign of the contrast. The first version was a neuromorphic system with a frame-based output. A
further version of the retina [21]-[22] had AER output and was an event-driven system. Nevertheless, it also had some drawbacks like high mismatch and bandwidth consumption when there was
no contrast.
After this approach, we have to highlight the work of Barbaro and Ruedi et al. [15]-[16].
They designed two devices. The first one was based on current mode and the second one was a
improved version based on voltage mode. It was close to frame-devices because events were
transmitted after a global reset. However, there was no information loss because information were
immediately available after resetting the device and spikes were sent in order. The device had
high dynamic range (120dB) and very low contrast mismatch (2%). It could detect spatial contrast
and gradient direction. The main limitations of this retina were the power consumption (300mW)
and temporal redundancy. The device was not able to detect the temporal contrast and its temporal
74
resolution was limited by the frame rate. There is also an improved version of this retina [17], it
has higher resolution (320 x 240), 132dB of dynamic range, and very low contrast mismatch
(0.5%).
The contrast retina presented in this thesis [23] and explained in chapter 4 is an improved
version of Boahen’s retina. It solves the main limitations of Boahen’s approach and has some
additional features like calibration, thresholding, time-to-first spike mode and bipolar output.
1.4.3 Temporal Contrast Extraction
Motion provides important information. This information is useful for the perception of relative depths, 3D structures, and for grouping objects that we see. The motion pathway is also
independent of other visual field pathways and plays a very important role.
The first attempt to implement an AER temporal contrast sensor was done by Mallik et al. in
2005 [7]-[8] and was based on the prior work of V. Gruev et al. [50] describing a pipelined temporal difference imager. Some previous reported retinae were also able to detect temporal contrast,
but they had low bandwidth and they were not specific devices for temporal contrast detection.
Mallik’s retina modified the traditional active pixel sensor (APS) to detect a quantize absolute
change in illumination. It was a frame-based device. Temporal contrast was computed from the
absolute difference between frames, so events times were discrete and not continuous. It had the
typical advantages of APS sensors: good fill factor (17%), pixels with reduced size and very low
fixed pattern noise (0.5%). Its bandwidth was limited by the frame rate and its dynamic range was
relatively low (51dB) compared to other AER sensors.
The first temporal contrast AER sensor (also known as the optical transient sensor) was
designed and proposed by Jörg Kramer [61]-[62]. It was a combination of Tobi Delbück’s adaptive photoreceptor [63] with a rectifying and thresholding differentiating element [64]. The outstanding sensor proposed by Lichsteiner and Delbrück [10]-[9] was based on the Kramer’s optical
transient sensor and has high dynamic range (120dB), low latency (15 µs ), good fill factor
(8.1%), and low noise (2.1% contrast). That sensor calculates the temporal derivative of the relative illumination. It is an asynchronous sensor with 128 x 128 pixels. Its bandwidth was limited
by the photoreceptor bandwidth.
In chapter 5, we present a new AER temporal sensor based on that work. It has lower
latency, higher bandwidth and better fill factor. This sensor is specially useful for high speed
applications and surveillance. It uses amplifiers at the photoreptor stage to improve the sensitivity
and the speed response. It also has a more efficient scheme of arbitration [47] that makes it faster.
Pixel size has been reduced 50%.
1.4.4 Time-to-First Spike Imagers
Vision Systems are not based on frames, however the data that arrive immediately after the
end of a rapid eye movement could be considered as a frame. Sometimes, it is useful to stop the
data transmission and then restart it without data loss (bandwidth consumption control). The first
time-to-first spike sensors [7]-[12] are based on S. Thorpe [11] ways of processing the visual
information like it is processed in the human visual system. After a global reset, the most active
pixels (pixels with more relevant information) will spike first. After a certain period of time, we
will only receive redundant information and noise. If we stop the data transmission, we will
improve the quality of the image and the bandwidth consumption will be reduced.
The new AER spatial contrast extraction retina presented in this thesis has an optional timeto-first spike mode that allows to combine the advantages of the frame-based processing in an
75
AER system: fast response and adjustable frame time. The sensor has a global reset. After presetting the retina, it will stop sending information to the AER bus. Thus, frame time can be set autonomically by adjusting the global preset period.
1.4.5 Luminance Sensors
These sensors [19]-[20] (also called octopus retinae) detect the light intensity and generate a
train of spikes with a frequency proportional to the local luminance. Although these sensors just
sense the illumination level, they have some advantages over frame-based devices, like continuous detection. However, spike frequency depends on the light and dynamic range is reduced in
comparison to other kinds of AER retinae. For low illumination levels, the output frequency can
be very slow.
1.5 New Developed AER Sensors
The goal of this work was to develop two different vision sensors that can detect spatial and
temporal contrast. Both sensors improve the state of the art and are insensitive to illumination
conditions and propose new ideas that can be used in further AER sensors.
Mismatch is a problem that has plagued prior AER retinae. For this reason, we have also
proposed a new calibration mechanism, that has been validated experimentally and used in the
design of the contrast retina.
The first sensor (described in chapter 4) can detect spatial contrast. It is a 32 x 32 pixels prototype that has some advantages over prior designs. It is insensitive to illumination conditions, it
has a new thresholding mechanism, an optional time-to-first spike mode and calibration.
One of the main disadvantages of prior AER contrast retinae was high mismatch. For this
reason, we developed a new compact calibration system (described in chapter 3) for large neuromorphic arrays [13]. This calibration technique has been successfully used to calibrate the mismatch of the contrast retina and offers higher resolutions that other calibrations techniques used in
prior designs [14].
The second sensor (described in chapter 5) detects temporal contrast. It is a 128 x 128 pixels
retina with high bandwidth and very low latency. It is specially useful for high speed applications
and surveillance.
The thesis has five chapters. Chapter 2 presents the Stochastic I-Pot, a compact and powerful circuitry to polarize and test circuits. Chapter 3 describes the new proposed compact calibration system (Newcalib) for neuromorphic arrays. In chapter 4 we present the new spatial contrast
AER retina. We explain thoughtfully the pixel desing and we show experimental results. In chapter 4, we present the temporal contrast AER retina. Explanations about the pixel design and experimental results are also provided. Finally, in chapter 6, we draw some conclusions and possible
future work is proposed.
Table 1.1 [57] summarizes the more important functionalities and performance figures of
prior AER sensors. Three types of functionalities are considered: sensing pixel luminance, sensing pixel temporal contrast, and sensing pixel spatial contrast with respect a given neighborghood.
For (spike) signal coding, three methods are used: signal to frequency (rate) coding, signal to
number of events (NE) coding, and signal to time-to-first spike (TFS) coding. When using ratecoding (as in [21]-[22]-[14]-[23]), a current that carries the information of interest (luminance,
contrast) is fed to an integrate-and-fire circuit whose spike frequency is controlled by the current.
For NE coding (as in [7]-[9]-[24]), every time the light sensed by a pixel changes by a relative
76
amount, a new spike is generated. In TFS coding (as in [20]-[16]-[23]), the information signal is
also fed to an integrate-and-fire circuit, but the integrators are (periodically and) globally reset
and only fire one spike between consecutive resets. This way, the information is coded as the time
between the global reset and pixel spike time. For a luminance retina [19]-[20], the photo current
is the one to be integrated, resulting in timings strongly dependent on ambient light. Consequently, the dynamic range of light sensing capability is directly transformed into the latency variation of the output. This is a severe practical restriction, labelled in Table 1.1 as “Light to Time
Restriction”. For contrast computations (either temporal or spatial), light difference is normalized
to average light, so that contrast is (by definition) independently of ambient light. Consequently,
these retinae should not suffer for “Light to Time Restriction”. This is the case for all contrast retinae in Table 1.1, except for [19]-[20]-[7]-[16]. [19] and [20] are luminance retinas. In [16], for
each frame there are two separate steps in time. The first one uses a Light to Time integration
(which lasts between 0.5 µs - 150ms depending on the ambient light) to obtain a voltage representation of pixel contrast. The second step transforms these voltages into a TFS representation
requiring an ambient-light-independence time of about 2ms. In [7], pixels use an modified active
pixel sensor (APS) and pixels can only detect absolute changes in illumination, so there is a also a
light to time restriction. Actually, the dynamic range is reduced in comparison to the other
devices.
Finally, the features and functionalities of our two new sensors are presented. The most significant features of each one have been highlighted.
77
Table 1.1: Comparison among AER vision sensor devices [57]-[23].
Mallik05
[7]
Licht07[9]
Posch10
[24]
Zagh04
[21]-[22]
Ruedi09
[16]
Costas07
[14]
Spatial
Contrast
Retina
[23]
Temporal
Contrast
Retina
Cul03[19]
Chen07[20]
Functionality
Luminance to
Frequency
Luminance
to TFS
Temporal
Contrast
to Number of
Events
Temporal
Contrast
to Number
of Events
Temporal
Contrast
to Number
of Events
Spatial
and Temporal
Contrast
to Frequency
Spatial
Contrast
Magnitude and
Direction
to TFS
Spatial
Contrast
to Frequency
Spatial
Contrast
to Frequency
or TFS
Temporal
Contrast to
Number of
Events
Light to
Time
Restriction
YES
YES
YES
NO
NO
NO
YES
NO
NO
NO
Latency
120 µs 125s
10 µs -1s
<5ms
15 µs @1
Klux chip
illumination
2Meps
3.5 µs @1
klux
Not
Reported
2ms150ms
Not
Reported
0.1ms10ms
3.5 µs @
25Klux
Dynamic
Range
120dB
>100dB
51dB
120dB
125dB
@30FPS
50dB
110dB
100dB
>100dB
>100dB
Spatial
Contrast
Computation
N/A
N/A
N/A
N/A
N/A
difusive
grid
neighbourhood
4 nearest
pixels
(up, right,
left, bottom)
diffusive grid
neighbourhood
(adjustable up
to 10
pixels)
diffusive grid
neighbourhood
(adjustable up
to 10
pixels)
N/A
Temporal
Contrast
Computation
N/A
N/A
Temporal
frame-difference
intensity
charge
detection
Temporal
derivative
of the relative
luminance
Temporal
derivative
of the relative
luminance
Temporal
derivative of the
relative
luminance
N/A
N/A
N/A
Temporal
derivative
of the relative luminance
FPN
4%
4.6%
0.5% of
full scale,
2.1% TD
charge
2.5%
SNRtyp
>56dB
1-2dec
1.7%
6.6%
0.6%
2.8%
Fill Factor
14%
33%
17%
8.1%
10%
14%
20%
3%
2%
8%
Pixel Size
32x30
17x17
25x25
40x40
30x30
34x40
14x14
58x56
80x80
35x35
AMIS
0.5 µm
3M 2P
0.35 µm
4M 2P
180nm
4M 2P
MIM
0.35 µm
2M 2P
180nm
1P6M
0.35 µm
4M 2P
0.35 µm
4M 2P
0.35 µm
4M 2P
[email protected]
5V
(50fps)
24mW
50-175
mW
63mW
300mW
33 µW 10mW
0.666.6mW
132-185
mW
µm
Fabrication Process
0.6 µm
3M CMOS
Power
3-71mW
0.35 µm
5M 1P
CMOS
N/A
78
CHAPTER 2
Biasing for Neuromorphic Arrays: The I-Pot System1
2.1 Introduction
Analog circuits require, in general, a set of bias currents which are usually provided by a
current reference circuit allocated somewhere at the periphery of the chip together with some current scaling and distribution circuitry [29], [30]-[31]-[32]. Many times, circuit designers would
like to be able to fine tune some of the designed bias currents in order to compensate for process
variations and components mismatches. In other occasions, when experimenting with new circuits, it is desirable to have large degrees of freedom to play with all available bias currents and
test the circuits for wide ranges of operating conditions. Under these situations, the safest solution
is to provide one external pin for each bias current, making accessible a transistor gate. This
allows to connect all available gates to external potentiometers (or a pot-box) so that one can
experiment freely in the lab with all combinations of biases. In practice, this approach is limited to
a reduced number of freely adjustable biases, since the number of pins of a chip cannot grow arbitrarily. To overcome this problem, Hasler et al. introduced the E-Pot circuit block [33], which
exploited floating gate and tunneling/injection circuit techniques to program non-erasable analog
voltages onto gates of biasing transistors. However, floating gate and tunneling/injection techniques are still complicated to use successfully in standard CMOS processes.
In order to generate programable bias currents we introduce an alternative approach, based
on the use of what we call the ‘Stochastic I-Pot’ concept [37]. The ‘Stochastic I-Pot’ is a digitally
programmable current source which, from a reference current, can provide any desired current
with high precision and down to pico-amperes. Each I-Pot cell includes a digital register to select
a current range, a current value for this range, and a sign. I-Pot cells can be chained so that any
arbitrary number of current biases can be generated and independently programmed. The number
of external pins that a chip needs for characterizing and programming the chain of I-Pots is three,
independent of the number of I-Pots assembled. Consequently, designers can include any arbitrary
number of programmable current biases, with the only restriction of area consumption. In our par-
1. This chapter of the thesis is based on a previous reported paper ([37]).
79
Chapter 2: The I-Pot System
ticular implementation, in a 0.35 µm process, the I-Pot cell area is 130µm × 68µm , one third the
area the pad it replaces.
We used the Stochastic I-Pot to generate all the bias currents that we needed to provide to
the contrast retinae.
2.2 I-pots Circuit Architecture
The ‘Stochastic I-Pot’ exploits the use of current mode ladder structures applied to MOS
transistors [34]-[28]. A generic MOS ladder structure, configured as current source, is shown in
Figure 2.1. All transistors are proportional to a unit size ratio W/L by either a factor 1, N-1, or N/
(N-1). This way, branch currents Ii have an attenuation ratio of N from each branch to the next
one. In the ‘Stochastic I-Pot’ circuit, we use two of these ladder structures (as we can see in the
circuit diagram of complete I-Pot cell in Figure 2.2)
The first ladder structure (Range Selector Ladder), with attenuation ratio around N=10, selects
an operating current range. This range ladder has 6 output branches, so that the output current can
be selected between the input reference current IREF and around IREF/106.
For the second ladder structure (Stochastic Ladder), we use an attenuation ratio of N = 2 .
This allows for selecting any binary combination of current branches, in the same way a current
DAC would do. However, we do not want to have a high precision (like 8 bit) current DAC within
each I-Pot, because they would require extremely large transistor sizes and would most probably
not provide such precision for very small currents down to pico-amperes.
In our approach we use ladder structures with attenuation ratio N = 2 , with a large number
of branches, but with small transistor sizes so that we intentionally provide large mismatches
between the current branches. By having a large number of branches, each with large mismatch,
we achieve a good coverage of possible output current combinations. Consider a ladder with input
current 300pA , attenuation ratio N = 2 , and 8 branches. The unit transistor size is
W ⁄ L = 1 ⁄ 0.7 µm . For a transistor of this size, fabricated in the AMS 0.35 µ m CMOS process,
and driving a current of ∼ 700 pA , results in a current mismatch standard deviation of around
σ ≈ 35% , as can be seen by interpolating the measurements shown in Figure 2.3, [28]-[35]. Figure 2.4(a) shows the output current obtained as function of the 8-bit digital word wval that controls
I REF
N−1
N−1 M
a
I
I m = REF
m−1
I
I2= REF
I 1 =I REF
N
Ma
Ma
N
Mb
I REF
N
m−1
(N−1)
N−1
N−1
N−1
Ma
N
I m+1 =
Ma
N
Mb
1
Mf
N−1
N−1
Fig 2.1: Generic ladder structure with N branches. The value of the current in
each branch is indicated.
80
Chapter 2: The I-Pot System
Range
Selector
Ladder (N=10)
IREF
IREFcopy
Current
Mirror
sw1
sw2
sw3
Irange
Io
VG
Ia
To external
I−test
(N=2)
Data−in
Stochastic
Ladder
wval
wrange
wsign wtest
Data−out
Clock
to j−th
bias
Fig 2.2: Circuit diagram of complete I-Pot cell.
the combination of branches as is shown in Figure 2.2. As can be seen, this characteristic differs
dramatically from a conventional stair-case that a high-precision 8-bit DAC would provide. However, suppose we introduce a look-up table between the digital word we provide and the one physically applied to the ladder structure, so that the output currents become ordered. The result is
shown in Figure 2.4(b) for the same ladder and bias reference current. Now we see a monotonic
increasing dependence between the digital control word word and the output current. It does not
matter that this relationship is not perfectly linear. Our objective is simply to provide a bias current as close as possible to a desired value. Such objective is limited by the intervals between consecutive current values in Figure 2.4(b). To characterize these intervals, we show in Figure 2.4(c)
the relative difference between the consecutive values in Figure 2.4(b),
In – In + 1
∆ rel = 2 --------------------In + In + 1
(2.1)
As can be seen, we have errors below 10% for the whole range except the first 1/20 of the
range, and below 1% for last 1/4 of the range. But this was a ‘lucky’ particular case. It is perfectly
possible to find situations where the mismatch plays against us and we obtain an unfortunate
extremely large maximum gap. This is for example, the case illustrated in Figure 2.4(d-f), for
another I-Pot, exactly equal to the one used for Figure 2.4(a-c). Such situations occur for example,
when the maximum branch current I 1 in Figure 2.1 results much larger than the sum of all the
others. This produces an extremely large gap in the center of the characteristic. In Figure 2.4(e)
the largest gap was of 92.3pA for a total range of 591.3pA . Consequently, as can be seen in
81
W=0.6µm
1
10
0
10
σ(∆IDS/IDS) (%)
σ(∆IDS/IDS) (%)
Chapter 2: The I-Pot System
−1
σ(∆IDS/IDS) (%)
−8
10
−7
10
−6
10
−5
10
−4
10
10
−3
10 (I )
DS
W=2.5µm
1
10
0
10
−8
10
−7
10
−6
10
−5
10
−4
10
−3
10 (I )
DS
W=5µm
1
10
0
10
−1
−9
10
σ(∆IDS/IDS) (%)
−9
10
−1
10
0
10
−1
−9
10
σ(∆IDS/IDS) (%)
10
W=1.2µm
1
10
−8
10
−7
10
−6
10
−5
10
−4
10
10
−3
10 (I )
DS
−9
10
−8
10
−7
10
−6
10
−5
10
−4
10
−3
10 (I )
DS
W=10µm
1
L=0.3µm
L=0.7µm
L=1.5µm
L=4µm
L=10µm
10
0
10
−1
10
−9
10
−8
10
−7
10
−6
10
−5
10
−4
10
−3
10 (I )
DS
Fig 2.3: Current mismatch standard deviation measurements, for NMOS transistors in a
0.35µm CMOS process [28]-[35]. Vertical axes represent standard deviation in %, and horizontal axes operating current. Measurements are taken for 30 different transistor sizes, by
sweeping width {10, 5, 2.5, 1.2, 0.6}µm and length {10, 4, 1.5, 0.7, 0.3}µm.
Figure 2.4(f), current values in the range between 250 to 350pA cannot be generated with a
precision better than 30%.
One can think of several solutions to solve this problem. After playing with a few of them and
testing them with statistical simulations, we found out that the most reliable solution is to
duplicate the output branches of the ladder with attenuation ratio N = 2 . This not only
guarantees there will be no large gaps between consecutive current values, but at the same time
reduces dramatically the value of the largest gap found, for the same transistor sizes and input
reference current.
Figure 2.4(g-i) shows the same situation for the circuit in Figure 2.4(a-c), but where now the
output branches are duplicated. Now there are 16 bits to select output branches combinations but
the maximum current gap in Figure 2.4(h) is now reduced to 0.20pA , excluding the first and last
100pA intervals of the range. The relative error is shown in Figure 2.4(i). As can be seen it is less
than 0.001 for currents between 137pA and 1214pA . Figure 2.5 shows the complete circuit
schematics of a ladder with attenuation ratio N = 2 and duplicated output branches.
Looking at the complete circuit diagram of a stochastic I-Pot cell in Figure 2.2. Each
stochastic I-Pot receives a copy I REFcopy of a common reference current I REF , which is the input
to a PMOS N = 10 range ladder structure with 6 output branches. The common reference
current can be generated by any bandgap type circuit [36] with temperature compensation, or
provided off-chip through an extra pin. The digital word wrange selects just one of the range ladder
outputs (not a combination of them). This current I range sets the coarse mode range of the I-Pot
82
(a)
500
400
300
200
100
0
10
(b)
−1
(c)
10
∆rel
500
400
300
200
100
Current (pA)
Current (pA)
Chapter 2: The I-Pot System
−2
10
−3
10
−4
10
50
100
150
200
250
50
100
w
150
200
250
1
w
val
ord
2
3
4
Current (x100 pA)
5
2
3
4
Current (x100 pA)
5
50
100
150
200
500
400
300
200
100
250
10
(e)
−1
−3
50
100
150
200
250
1
w
val
ord
1.2
1
0.8
0.6
0.4
0.2
0
10
(h)
−1
10
−2
∆rel
(g)
Current (nA)
Current (nA)
−2
10
10
w
1.2
1
0.8
0.6
0.4
0.2
(f)
10
∆rel
(d)
Current (pA)
Current (pA)
0
500
400
300
200
100
(i)
10
−3
10
−4
10
−5
10
1
2
3
4
wval
5
6
x 10
1
2
3
4
word
4
5
6
x 10
2
4
4
6
8 10
Current (x100 pA)
12
Fig 2.4: (a-c) Illustration of mismatch effects in a MOS ladder with ratio N=2. (d-f)
Same, but for an ‘unlucky’ ladder example. (g-i) Illustration of a ladder structure with
duplicated output branches (a,d,g) I-Pot output currents versus digital word wval,
(b,e,h) same output currents after ordering versus word, (c,f,i) relative difference
between consecutive values.
output current. Current I range is now fed to the input of an NMOS N = 2 ladder structure with
duplicated output branches. Let us call this ladder a “stochastic ladder”. In our particular case we
implement 16 ( 2 × 8 ) duplicated branches. The particular combination of output current
branches is controlled by the digital word in register wval. The output of this ladder I a can be
optionally sign inverted by transmission gates sw1 – sw3 , controlled by the state of an extra
register w sign, which inserts or not a PMOS current mirror in the output current path. Finally, the
signed output current I o is directed to either its destination bias point, or to a chip output pin I test
for characterization purposes, depending on the state of register w test . All registers, wrange, wval,
w sign and w test are shift registers, connected in series, and clocked by the same clock signal. I-Pot
cells can be chained directly in series by chaining their shift registers sharing all the same clock
signal. All I-Pot cells share also the same gate line V G and test line I test .
83
Chapter 2: The I-Pot System
b
b
15 14
b
13
b
12
b
b
11 10
b
3
b
2
b
0
b
1
I
out
I range
I
dump
W/L
W/2L
a
I1
W/2L
a
I2
W/L
W/2L
W/2L
a
I3
W/2L
W/2L
W/L
W/2L
W/2L
W/L
W/L
b
I1
a
I8
a
I7
b
I2
W/2L
b
I3
W/2L
b
I7
W/2L
b
I8
W/2L
W/L
W/L
Fig 2.5: Circuit schematic of ladder structure with attenuation ratio N=2 and duplicated output branches.
The main drawback of the present I-Pot approach is that each I-Pot of each fabricated chip
needs to be characterized individually. The good news are that this is quite easy to do by using a
host computer that loads the chained shift registers, while at the same time we require the use of
just one single external current metering instrument.
The procedure for characterizing the I-Pots of a chip is as follows:
1.- Each I-Pot has to be characterized individually. Consequently, the w test bit of only one single
I-Pot has to be set to ‘active’. All others must be disabled. This way, only one single I-Pot output is connected to external line I test.
2.- Sweep the two signs for the active I-Pot.
3.- For the selected I-Pot and sign, sweep all current ranges through digital word wrange.
4.- For the selected wrange, sweep all 20 output current branches, measure the selected branch
with the external current meter through pin I test, and store it in a file in the computer.
After completing the measurements for one I-Pot, we will have stored in the computer a total of 2
signs x 6 ranges x 16 branches = 192 current values. For each sign and range, we can now produce
16
all 2 possible combinations, order them, and find the maximum gap. What now remains to do,
is writing a computer program that given a desired I-Pot value will return the optimum wval and
wrange that gives the minimum error. The total amount of time required to caracterize one individual I-Pot was about 1 hour.
84
Chapter 2: The I-Pot System
−10
−9
x 10
−8
x 10
x 10
3
3
2.5
2.5
2
2
1.5
1.5
4
3
1
1
0.5
0.5
2
1
0
2
4
0
6
2
4
4
2
4
4
x 10
−7
6
4
x 10
x 10
−5
x 10
−4
x 10
x 10
1.2
6
1.5
1
5
4
0.8
3
0.6
2
0.4
1
0.2
0
0
6
2
4
0
6
1
0.5
2
4
4
0
6
2
4
x 10
x 10
4
6
4
x 10
Fig 2.6: Measured currents of one of the fabricated I-Pots for all six current ranges. Continuos lines show the same values after ordering. Vertical axes are measured currents, horizontal axes are code values for the 216=65536 currents of each range.
2.3 Experimental Results
Here, we present some experimental results after characterizing the I-Pots. I-Pots were fabricated in the AMS 0.35 µ m CMOS process. The stochastic ladder uses 16 duplicated output
branches, and the unit transistor of the ladder structure has a size of W = 1µm , L = 0.7µm .
This ladder was intentionally made with a small unit transistor to increase mismatch, and therefore increase its stochasticity. The range ladder used 6 output branches, and the input current to
the range ladder was set to 100µA . The range ladder was designed according to Figure 2.1 with
an attenuation ratio of N = 11 , approximately.
After measuring one of the fabricated I-Pots we obtain the currents shown in Figure 2.6.
Each of the six subgraphs corresponds to one of the available ranges provided by the range ladder.
In each subgraph we have added the ordered version of the measured current values. This ordered
version is drawn with the continuous line. The minimum and maximum current values provided
by each range are as follows. First range [0.4mA, 176mA], second range [30nA, 12.7mA], third
range [1.9nA, 673nA], fourth range [126pA, 44.0nA], fifth range [9.9pA, 3.28nA], and sixth range
[1.2pA, 331pA].
Figure 2.7 shows these values, expressed in bits2 as function of currents. Striation effects
can be observed, specially as current decreases, due to quantization effects in the data acquisition
instrument.
2.
1⁄2
n bits
= ∆ rel ⇔ n bits = – log 2 ( ∆ rel ) = – ( ln ∆ rel ) ⁄ ( ln 2 )
85
Chapter 2: The I-Pot System
14
14
14
12
12
12
10
10
10
8
8
8
6
6
6
4
4
4
2
2
2
0
0
1
2
0
3
0
1
2
0
3
−10
−8
14
12
12
12
10
10
10
8
8
8
6
6
6
4
4
4
2
2
2
2
4
0
6
4
x 10
14
0
2
x 10
14
0
0
−9
x 10
0
5
−7
0
10
−6
x 10
x 10
0
5
10
15
−5
x 10
Fig 2.7: Computed relative increments between ordered consecutive values expresed in
bits. Vertical axes are resulting bits, while horizontal axes are measured currents.
We can see that the maximum resolution is obtained for the central parts of the ranges,
reaching values as high as 13-bits. On the other hand, for the external parts of the ranges the resolution decreases dramatically down to values as low as 1-bit. This is because in the central part of
the ranges there is a higher density of redundant values than on the extremes of the ranges. However, if the ranges overlap, we can increase the density of redundant values and improve the resolution. The procedure is as follows. Let us take all the measured values of all six ranges and order
them as one unique set of current values. Each current value is uniquely defined by its range and
its 16 bit word within that range. Let us now compute the difference of consecutive values, as
defined by eq. (2.1) on page 63, and express them in bits The result is shown in Figure 2.8. We
can still see very well the regions of the six ranges with their respective maximum resolution central parts of up to 13-bits. However, the resolution of the extreme regions of the ranges has
improved to values of above 8-bits. Neglecting the two extreme regions of the whole merged six
ranges, the worst resolution is obtained for currents around 0.7mA, yielding a resolution of 8.05bits. The first current value showing a resolution above 8 bits is 19.6pA, and the largest one
176mA.
2.3.1 Temperature Effects
Ideally speaking, since the I-Pot operation is based on current splitting techniques, the effect
of temperature should be minor as long as the current division ratios between ladder branches
remain unaltered. Unfortunately, this is not completely true, and those ratios do depend slightly on
temperature. We have observed that those ratios can vary between 0.1% and around 1% for tem-
86
Chapter 2: The I-Pot System
13
12
Precision (bits)
11
10
9
8
7
6
5
4
−11
10
−10
10
−9
10
−8
10
−7
10
−6
10
−5
10
−4
10
Measured Current (A)
Fig 2.8: Computed relative increments, expressed in bits, when considering a measured values if all six ranges as a unique set. Vertical axis represents resulting bits, while horizontal
axis represents absolute measured current.
perature variations between 0º-40ºC for currents above tens of nano-amps. For lower currents the
effect of temperature is more drastic, since leakage currents increase rapidly with temperature.
Figure 2.9 shows the error (expressed in bits) between the I-Pot current expected when characterized at 20ºC, and the one obtained when doing new measurements, sweeping temperature
between 0ºC and 40ºC. Each data point is obtained by taking the maximum error within 1/30-th of
the 7-decade current range.
2.4 Conclusions
In this chapter, we have presented the Stochastic I-Pot. It is a versatile and powerful circuit
for generating digitally controlled precise bias currents has been presented. 8-bit precision has
been verified from currents as low as 19.6pA up to values of 176mA (almost 7 decades). Temperature degrades precision gracefully, and is more severe for smaller currents. This circuit is specially handy for experimenting with current-mode circuits operating in weak inversion, where
mismatch is high and operating range extends over several decades. It is also very useful for
experimenting with new circuits, where it might be desirable to include a large number of finetunable current biases for trimming gains and offsets. The only drawback is that each I-pot needs
to be characterized individually using an external off-chip current metering instrument.
87
Chapter 2: The I-Pot System
12
10
Precision (bits)
8
6
0ºC
5ºC
10ºC
15ºC
20ºC
25ºC
30ºC
35ºC
40ºC
4
2
0 −11
10
−10
10
−9
10
−8
10
−7
10
Current (A)
−6
10
−5
10
−4
10
Fig 2.9:Effect of temperature on precision.
I-pots were successfully used to test the two retinae presented in the thesis. In both cases, we
set I ref = 17µA to achieve more precision with lower currents values.
88
CHAPTER 3
The New Calibration System1
3.1 Introduction
Over the last 20 years, a vast amount of neuromorphic VLSI systems have been reported ([6][22], just to mention a few) which usually consist of large arrays of special processing pixels.
Since pixels have to be of reduced size and power consumption, analog design techniques are
used with transistors of small size operating with nano amps or less. This yields necessarily high
mismatch. Although reported neuromorphic VLSI systems have revealed interesting, powerful,
and fast information sensing and processing capabilities, they still have not evolved clearly to
specific marketable products. One of the main reasons for this is the unavoidable excessive
mismatch that plagues most of the so far reported neuromorphic chips [6]-[22].
To keep mismatch low without increasing transistor sizes nor operating currents, the only
known solution is calibration. Some researchers have reported calibration techniques based on
floating-gate MOS transistors [25]-[26] in standard CMOS processes. However, these techniques
require a special know-how, which makes difficult the path towards marketable products.
Recently, some neuromorphic systems with in-pixel RAM based calibration techniques have been
reported [14]-[27], which exploit the use of compact current DACs made with calibrable MOS
ladder structures [28]. The drawback of this approach is that it uses a one-point calibration principle, which limits the final precision to 3-bit for practical transistor currents and sizes. In this thesis, we present another way of implementing a digitally adjustable MOS, [13], much more
compact than a MOS ladder structure, which makes viable a multi-point calibration. Also, we
introduce some extra translinear circuitry which enables to sweep operating currents of the calibrated current sources without requiring recalibration. This way, circuit precision degrades
smoothly when changing operating currents.
By using a calibration system, the performance of a system can be improved. However, adding some calibration circuitry has some drawbacks. Obviously, the circuit has to been calibrated at
least once. This implies time and it is specially important when electronics circuits are fabricated
massively. Calibration time has to be reduced as much as possible. Another issue is area. Calibra1. This chapter of the thesis is based on the previous reported paper, ([13]).
89
Chapter 3: The New Calibration System
b
b
2
N−1
b0
b
1
D
V
dummy
W/L
W/L
G
W/L
W/L
W/L
(a)
S
W/L
G
W/L W/L
W/L
W/L W/L
W/L
W/L
D
W/L
W g( w )
cal
L
(b)
wcal
S
Fig 3.1: Prior reported digitally controlled width MOS transistor. (a) Circuit schematics
and (b) symbol
tion circuitry inside pixels has to be simple and reduced. If not, we will degrade the fill factor and
we will increase the total area of the circuit considerably.
Another point to consider is how to estimate the maximum admissible mismatch. This is not
trivial, because the values of some currents (i.e. diode photo current chapters) are not always
known and can vary significantly. Once the mismatch is known, the calibration system can be
designed and we have to decide if it makes sense to add it to the pixel. Sometimes, it could be better to increase the transistors area or to use higher bias currents to decrease mismatch effects.
When we design a calibration system for a large neuromorphic array, the number of transistors has to be as low as possible. We cannot try to oversize the system. It is also desirable to
implement calibration systems whose effects do not degrade too fast when bias currents are scaled
or there is a change in illumination. We would like to calibrate the system in certain average conditions and then preserve the calibration when bias are scaled or there are different settings in the
visual scene. It would be also desirable to do the calibration on chip without external circuitry just
one time.
One of the targets of the thesis was to design a contrast sensitive retina based on the previous
work of Boahen. After studying and simulating the circuit, we checked that one of the main drawbacks was mismatch. For this reason, we tried to develop new calibration techniques that could be
used in generic neuromorphic VLSI systems. Not only does calibration improve mismatch, it also
makes possible to work with lower bias currents and reduced consumption or to have a bipolar
output (as we will discuss in chapter 4) in large arrays of pixels.
In this chapter, we will explain the new calibration system and its more important features.
We will present some experimental results. Its specific use in the contrast retina will be discussed
in the next chapter.
3.2 MOS Transistor with Digitally Adjustable Length
Previously reported in-pixel RAM based calibration circuits [14]-[27] were based on the use
of MOS ladder structures [28]. Figure 3.1 shows the circuit of an equivalent N-bit digitally
adjustable MOS transistor. This circuit is composed of 3N + 1 unit transistors of size W/L and N
switches. Because of the current splitting principle in MOS ladder structures [32] each vertical
branch drives a current half of the previous one, except for the last one. The circuit in Figure
3.1(a) between terminals D, G and S behaves (at DC) identically to a MOS transistor, as in Figure
90
Chapter 3: The New Calibration System
VDD
D
bN−1
IREF
S N−1
G
Ical
W
L
b1
S1
W
L g( wcal )
b0
S0
(a)
W
L
wcal
S
Ib
(b)
Fig 3.2: (a) Schematics of the digitally controlled length MOS transistor. (b) Application to
a calibration current source.
3.1(b), of length L and whose equivalent width is digitally adjustable through the N-bit digital
word
w cal = { b N – 1 b N – 2 …b 2 b 1 b 0 } .
The
equivalent
width
is
given
by
W eq = ( W ⁄ 2 ) × g ( w cal ) , where
N–1
g ( w cal ) =
bi
∑ ----------------N–1–i
2
(3.1)
i=0
Digital word w cal is stored on in-pixel static RAM cells at startup. The optimum calibration
words are obtained from a calibration procedure followed the first time the chip is used [14]-[27].
From eq. (3.1) we can see that the equivalent width can be adjusted between 0 and
N–1
N–1
.
2 – (1 ⁄ 2
) in steps of 1 ⁄ 2
To develop calibration circuitry for large neuromorphic arrays, we developed a new
approach to digitally adjust the size of a MOS transistor using a more compact circuitry [13]. Figure 3.2(a) shows the schematics of the proposed circuit. There are N transistor segments between
terminals D and S. Each segment is either enabled by connecting its gate to terminal G, or disabled
by connecting its gate to VDD. Transistor sizes can be for example S N – 1 = W ⁄ L ,
S N – 2 = 2W ⁄ L , ... and
S0 = ( 2
N–1
)W ⁄ L . This can be implemented physically by using
unit transistors of size W ⁄ L (one for S N – 1, two in parallel for S N – 2 , ... 2
N–1
in parallel for
N–i–1
S 0 ). This way, each segment would be equivalent to a transistor of size S i = W ⁄ ( L ⁄ 2
) .
Consequently, the digitally adjustable transistor in Figure 3.2(a) would be equivalent to one of
N–1
width W and digitally adjustable length L eq = L × g ( w cal ) = L ×
bi
----------------∑ 2N – 1 – -i . This transistor
i=0
can be used as part of a current mirror1, as shown in Figure 3.2(b), to provide a calibration current
91
Ical
Chapter 3: The New Calibration System
wcal
Fig 3.3: Montecarlo simulation (with 100 iterations) of the circuit in Figure 3.2(b), using a
4-bit digitally controlled length MOS.
N – 1 b

i

- + 1 . Figure 3.3 shows the simulated stairs
I cal = I REF × ( g ( w cal ) + 1 ) = I REF × ∑ ----------------N–1–i


2


i=0
of I cal as function of wcal (using a 4-bit digitally-controlled-length MOS) with I REF = 3nA ,
using unit MOS sizes of 1 µ /4 µ m, and models for a 0.35 µ m standard CMOS process. Figure
3.4(a) shows I cal as function of I REF before calibration, with w cal = 15 for each of the 100
simulated Montecarlo iterations. The mismatch at I REF = 3nA is ∆I cal ⁄ I REF = 110% and at
I REF = 1pA is 130%. Using the results in Figure 3.3 (for I REF = 3nA ), one can compute
for each Montecarlo iteration the optimum value of wcal for minimum spread at I cal . Once setting this optimum set of values for wcal, the resulting I cal as function of I REF is shown in Figure
3.4(b). Now the mismatch at I REF = 3nA has been reduced to 4% (4.6 bits).
1. Here we use a sub-pico-ampere current mirror topology [36], since we want to use eventually IREF values
down to the pico ampere range [27].
92
Chapter 3: The New Calibration System
Ical
(a)
Ical
Iref
(b)
Iref
Fig 3.4: Montecarlo simulation results for the circuit in Figure 3.2(b) when sweeping Iref.
(a) Before calibration with wcal=15 for all Montecarlo iterations. (b) After calibration with
optimum wcal for each iteration.
From a practical point of view, it is not efficient to follow the previous unit transistor based sizing
strategy. Note that transistor of size S N – 1 is the most critical for mismatch (since this segment
contributes the largest L). However, it uses only one unit transistor, while transistor of size S 0 is
N–1
the least critical for mismatch and uses 2
units. In practice it is more efficient to use one
single transistor for each MOS segment and adjust its size to have a similar effect. Furthermore,
from a statistical point of view, since mismatch plays an important role, the steps of the final stair
cases will not be all equal. The maximum step heights will limit the final calibration capability.
Therefore, it is important to minimize this maximum possible step height. To do this, the nominal
stair case should be designed with some intentional ‘down-steps’, so that when mismatch introduces random variations the extra redundancy compensates for eventual large up-steps. Figure
3.5, for example, shows Montecarlo simulation results of a 5-bit structure that uses one single
transistor per segment and has intentional down-steps. Simulated transistor sizes were {2/3, 2/1.8,
2/1.8, 2/1, 2/0.7}.
3.3 Translinear Circuits for Tuning
The calibration technique shown in Figure 3.2 requires to recalibrate all circuits when there
is a global change in the operating current I REF . In practice, it is desirable to allow a change in
the operating current I REF without requiring recalibration. Note that, all transistors introduce
mismatching and calibration compensates for the combination of all mismatches of all transistors.
The mismatch introduced by each transistor is dependent on its operation current and bias conditions. To have calibration less sensitive to bias conditions one should use topologies that change
93
Chapter 3: The New Calibration System
Ical
Ical
(a)
(b)
B
A
wcal
wcal
Fig 3.5: Example simulation of a 5-bit digitally controlled length MOS with one transistor
per segment and intentional down-steps. (a) Nominal mismatch-less simulation. (b) Monte
Carlo simulation with 100 iterations.
bias conditions for as few transistors as possible. To achieve this we introduce tunable translinear
circuits, which will allow us to keep fixed bias currents for some transistors, including the digitally-controlled-length ones. This is shown in Figure 3.6. The circuitry comprised by broken lines
is replicated once per pixel, but the rest is implemented only once at the periphery. Transistors M1
to M4 form a translinear loop, thus I oi = I 1 I 2 ⁄ I i3 . Local current I i3 is mirrored from the
peripheral global current I 3 , through a current mirror with a local digitally-controlled-length
MOS. To achieve a factor 2 calibration range, we include two transistors in series for this current
mirror output. One of fixed size W ⁄ 2L and the other calibrable. Consequently,
I i3 = I 3 ⁄ ( 2 + g ( w cali ) ) and
I oi
I1 I2
= --------- ( 2 + g ( w cal i ) )
I3
N–1
, with g ( w cali ) =
94
bi
----------------∑ 2N – 1 – -i
i=0
(3.2)
Chapter 3: The New Calibration System
wcali
I1
W
Lg(wcal )
i
M2
W
2L
W
L
M4
M1
M3
I2
Ii 3
Ioi
I3
Fig 3.6: Translinear circuit for tuning operating range of calibration circuit.
With this circuit, one can maintain (after calibration) constant currents I 3 (and I i3 ) and I 1 ,
while tuning I 2 globally to scale up or down all local currents I oi .
3.4 How to Optimize the Calibration Ranges.
For calibration, the goal is to find the optimal horizontal line that cuts through all stairs and
produces the minimum dispersion among all stairs. Note in Figure 3.5(b) points ‘A’ (top value of
left side) and ‘B’ (bottom of right side). If ‘B’ is below ‘A’ the maximum dispersion after calibration will be high, because there will be no horizontal line cutting all stairs. If ‘A’ is below ‘B’ it is
possible to find for each stair a value close enough to the desired horizontal line cutting all stairs.
For optimum calibration it is desired that ‘A’ be close to ‘B’, so that final calibration words may
spread over the whole range. The resulting relative position of points ‘A’ and ‘B’ depends on the
resulting mismatch distribution of the array and the resulting process corner of the sample. One
can design the nominal case to have ‘A’ as close as possible to ‘B’, but then many fabricated samples will result with ‘A’ higher than ‘B’ yielding poor calibration capability. On the other hand, if
one designs the nominal case for ‘A’ conservatively lower than ‘B’, then many samples will not
take advantage of all their bits for calibration, resulting in reduced calibration capability. Consequently, in practice, it will be desirable to be capable to adjust the relative positions of points ‘A’
and ‘B’ electronically. For this, we have implemented two different global optimization strategies.
In the first strategy, shown in Figure 3.7, two digitally-controlled-length transistors are used. One
of them is adjusted locally, as in Figure 3.6, but the other is adjusted globally. Thus all gates of its
transistor segments (see Figure 3.2(a)) are shared by all pixels and controlled from the periphery.
As a result,
I1 I2
I oi = --------- ( g ( w adj ) + g ( w cal i ) )
I3
(3.3)
Figure 3.8 shows the resulting simulated staircases for 3 different values of global control word
w adj . For one extreme ( w adj = 31 , as in Figure 3.8(a)) ‘A’ is above ‘B’, and the array has very
95
Chapter 3: The New Calibration System
wcal
i
I1
M2
M
W1 m
L1
wadj
M1
W1
L1g(wcal )
i
W1
L1g(wadj )
M3
I2
M4
Ioi
Ii 3
I3
Fig 3.7: First strategy for optimizing calibration range.
Ioi
Ioi
(a) wadj=31
Ioi
(b) wadj=20
(c) wadj=0
A
B
wcal
wcal
wcal
Fig 3.8: Simulation results for first strategy. (a) Simulated stairs for wadj=31, (b) for
wadj=16, (c) and wadj=0. Vertical scale is the same for the three graphs.
poor calibration capability. For the other extreme ( w adj = 0 , as in Figure 3.8(c)), ‘A’ is at the
bottom and the horizontal lines cut only a reduced range of the stairs, thus reducing significantly
the available number of bits for calibration. The optimum solution is an intermediate one, in this
case w adj = 20 as in Figure 3.8(b), which sets points ‘A’ and ‘B’ to be close. The optimum
value of w adj is sample dependent.
The second global optimization strategy is shown in Figure 3.9. Here the translinear circuit
has been replicated twice, so that there are two of such translinear circuits in parallel. One of them
uses local calibration through local digital control word w cali . The other is adjusted globally and
only the output transistor M x of its translinear set is replicated once per pixel. This allows for a
larger size of this transistor, and consequently less mismatch. The purpose of the locally cali-
96
Chapter 3: The New Calibration System
I’1
I’3
Mx
Ix
I’2
I1
wcal
i
M2
W1
L1
M4
M1
M3
Ii 3
I2
Ii 4
I3
Ioi
Fig 3.9: Second strategy for optimizing calibration range.
wcal
wcal
wcal
Fig 3.10: Simulation results for second strategy. (a) For all bias currents equal to 10nA, (b)
detail of the bottom calibrate subcircuit Ii4, and (c) results for turning bias currents Ii down
to 4.5nA.
brated translinear circuit is to compensate for the mismatch at M x . Figure 3.10 shows simulation
results for this circuit. In Figure 3.10(a) all peripheral bias currents I k and I k' were set to 10nA .
97
Chapter 3: The New Calibration System
The result is ‘A’ lower than ‘B’ and a reduced range for the calibration words. Figure 3.10(b)
shows the contribution of only the bottom locally adjustable subcircuit ( I i4 in Figure 3.9). Note
that for
w cali = 0
3.10(a), for w cali = 0
the bottom circuit does not add current to I oi . Consequently, in Figure
the mismatch is produced only by the upper M x transistors. Note that
this left part of the stairs will be fixed if peripheral currents I k' are maintained fixed. The tuning
strategy consists now in scaling peripheral currents I k until obtaining the optimum situation
shown in Figure 3.10(c). In this case, we have set all I k = 4.5nA . After finding the optimum
calibration words, the resulting operating point can be scaled by adjusting simultaneously only
peripheral currents I 2 and I 2' .
3.5 Experimental Results
In order to characterize the new calibration system, a test prototype microchip was fabricated
in a standard 0.35 µ m CMOS process. Twenty 5-bit current DACs were fabricated. Ten of them
used the first calibration range optimization strategy (Figure 3.7), and the other ten used the
second one (Figure 3.9).
Each of the first ten DACs uses five replicas of the circuit in Figure 3.7, one for each bit. The
nominal output currents of each ( I oi ) were adjusted to be binarily scaled. Consequently, at the
periphery, we need five groups of current sources {I 1 , I 2 , I 3 } and five groups of transistors {M 1 ,
M 2 , M m }, one for each bit. However, these five groups of peripheral current sources and
transistors are shared by all ten DACs.
Each of the second ten DACs uses five replicas of the circuit in Figure 3.9. Again, for each of
the ten DACs, the circuitry is replicated five times (one per bit), and the peripheral circuitry
(outside broken lines in Figure 3.9) is shared, per bit, by all ten DACs.
2
The area of the circuit layout inside broken lines is 18 × 14µm for Figure 3.7 and
2
17 × 14µm for Figure 3.9.
Figure 3.11(a) shows the experimentally measured output currents for ten replicas of the circuit in Figure 3.7, when setting w adj = 0 . Peripheral bias currents were made equal to I 1 = I 2 =
I 3 = 10nA , and all calibration words w cali (i = 1, ... 10) were swept simultaneously from 0 to 31.
After repeating this measurement for all possible w adj values, the optimum value for w adj
corresponds to the situation where the top left value is closest to the bottom right one. This case is
shown in Figure 3.11(b). At this point we can obtain the ten optimum calibration words w cali
that render the minimum variation. The maximum output current spread obtained under these circumstances is ∆I oi max = 0.57nA , which corresponds to 5.7%, at a nominal current of
I b = 10nA . If this were the current source controlled by the most significant bit of a current
DAC (with 20nA
maximum range), it would limit the DAC precision to
– ln ( ∆I oi max ⁄ 2I b ) ⁄ ln 2 = 5.13bits . To verify how calibration degrades when changing bias
conditions, we swept I 2 in Figure 3.7 between 100pA and 1µA . The maximum current spread
among all 10 calibrated current sources is shown in the trace with circles in Figure 3.12. The trace
with triangles are measurements obtained before calibration ( w cali = 0 , for all i). We can see the
10 samples maintain a precision of 4 bits for currents above 3nA. Horizontal axis is average of
98
Chapter 3: The New Calibration System
4.5
4
5
(b)
(a)
4
3.5
3
3
nA
nA
2.5
2
2
1.5
1
1
0.5
0
0
5
10
15
20
wcal
25
0
0
30
5
10
15
wcal20
25
30
Fig 3.11: Experimentally measured output currents for the circuit in Figure 3.7. (a) for
wadj=0, (b) for optimum wadj. The horizontal line in (b) is the target value, which is cut/
touched by all 10 traces.
I oi among all ten samples. We also show in Figure 3.12 the resulting precision after calibration
obtained through simulations, shown with crosses. Note that it is over optimistic, except for the
point at which calibration was done (10nA). The reason is that in this particular circuit (Figure
3.7) calibration degrades because of mismatch in MOS transistor slope factor. Mismatch in this
parameter is not modelled in our simulator.
In a similar way, Figure 3.13 shows the measured precision before and after calibration of ten
calibrable and tunable current sources that follow the approach depicted in Figure 3.9. Note that
now the mismatch before calibration is less than in Figure 3.12. This is because now the area used
by digitally-controlled-length transistor M b in Figure 3.7, is available for transistor M x in Figure
3.9, which can be made larger. With the structure of Figure 3.9 we obtain a much better precision
at the calibration point (8.30 bits at 10nA), but degrades rapidly, specially for high currents. The
precision after calibration obtained by simulation, is slightly pessimistic at the calibration point
(7.63 bits at 10nA), but it degrades optimistically as operating current departs from the calibration
point.
Figure 3.12 and Figure 3.13 show the matching precision among 10 current sources calibrated
at 10nA. Now we use 5 of these sources, calibrated at {10, 5, 2.5, 1.25, 0.625}nA, to build a 5-bit
current DAC. The matching precision obtained among the 10 fabricated DACs is shown in Figure
3.14 (for the tuning scheme of Figure 3.7) and in Figure 3.15 (for the tuning scheme of Figure
3.9). The DACs were calibrated at 16ºC, and the figures also illustrate the DACs behavior when
temperature is changed between 0ºC and 40ºC. We can see that the effect of temperature is not
severe for the lower current range, while for higher currents the DACs are almost insensitive to
temperature variations.
3.6 Discussion
In this chapter, we have described a new method to calibrate large neuromorphic arrays. The
approach is illustrated for current sources operating in the nano ampere range. Two tuning
schemes are proposed for sweeping the operating range over four decades. The first one achieves
99
Chapter 3: The New Calibration System
less precision at the calibration point but degrades more gracefully as operating current is
increased. The second one achieves higher precision at the calibration point but precision
degrades more as current increases. Test prototypes were fabricated and extensively tested and
characterized. As an example application, current DACs of 5-bit resolution and 20nA range have
been fabricated and characterized.
As it will be explained in the next chapter, the calibration procedure was used to reduce mismatch in the spatial contrast retina. In this case, the scheme of Figure 3.9 was chosen because we
were interested in getting the highest precision at a fixed calibrated point without large variations
of bias currents.
8
measured
simulated
uncalibrated
7
Precision (bits)
6
5
4
3
2
1
0
−10
10
−9
10
−8
10
Current (A)
−7
10
−6
10
Fig 3.12: Measured precision of calibrate and tunable current source with the approach of
Figure 3.7. Trace with circles: measured precision after calibration (with optimum w cal j for
each of the ten current sources). Current sources were calibrated at 10nA. Trace with triangles: measured precision before calibration ( w cal i = 0 for all current sources). Trace with
crosses: precision after calibration, obtained through simulations.
100
Chapter 3: The New Calibration System
9
measured
simulated
uncalibrated
8
Precision (bits)
7
6
5
4
3
2
1 −10
10
−9
10
−8
10
Current (A)
−7
10
−6
10
Fig 3.13: Measured precision of calibrable and tunable current source with the approach of
Figure 3.9. Trace with circles: measured precision after calibration. Current sources were
calibrated at 10nA. Trace with triangles: measured precision before calibration ( w cal i = 0
for all current sources). Trace with crosses: precision after calibration, obtained through
simulations.
101
Chapter 3: The New Calibration System
DAC Precision (bits)
5
4
3
2
1
0ºC
16ºC
30ºC
40ºC
−9
10
−8
−7
10
10
DAC Max Current (A)
−6
10
Fig 3.14: Measured Precision for the 5-bit DAC samples that use the first tuning strategy of
Figure 3.7. DACs were calibrated with MSB at 10nA and at 16ºC. After calibration, precision is characterized sweeping operating current for different temperatures.
102
Chapter 3: The New Calibration System
8
0ºC
16ºC
30ºC
40ºC
7
DAC Precision (bits)
6
5
4
3
2
1
−9
10
−8
−7
10
10
DAC Max Current (A)
−6
10
Fig 3.15: Measured precision for the ten 5-bit DAC samples that use the second tuning
strategy of Figure 3.9. DACs were calibrated with MSB at 10nA and 16ºC. After calibration, precision is characterized sweeping operating current for different temperatures.
103
CHAPTER 4
The AER Spatial Contrast Sensor1
4.1 Introduction
Visual sensors (retinae) are among the first AER modules to be reported since the introduction of the technology [1]-[2]. Spatial contrast AER retinae are of special interest since they provided highly compressed data flow without reducing the relevant information required for
performing recognition. As we explained in chapter 1, a variety of AER visual sensors can be
found in the literature, such as simple luminance to frequency transformation sensors [19], Timeto-First-Spike (TFS) coding sensors [20]-[16], foveated sensors [39]-[40], more elaborate transient detectors [9]-[24]-[41], motion sensing and computation systems [42]-[43], and spatial and
temporal filtering sensors that adapt to illumination and spatio-temporal contrast [21]-[22].
Spike based visual sensors can code their output signals using rate coding or TFS coding.
When using rate coding, each pixel is autonomous and continuously generates spikes at a frequency proportional to the signal to transmit (such as luminance or contrast). Under such circumstances, there are no video frames, so that sensing and processing is continuous and frame-free.
When using TFS conding, a global system-wide reset is provided and each pixel encodes its signal by the time between this reset and the time of the only spike it generates. Sensing and processing is frame-constraint. However, TFS is a highly compress coding scheme (each pixel generates
at the most one spike per frame) and frame time can be dynamically adjusted to an optimum minimum by subsequent processing stages. TFS and subsequent ideas were originally proposed by
Thorpe based on neuro physiological and psycophysical experiments [11], and have evolved to
practical very high speed image processing software techniques [44].
Computing contrast on the focal plane significantly reduces data flow, while relevant information for shape and object recognition is preserved.
Previously reported plain spatial contrast retinae [45]-[4]-[14] compute a contrast current
per pixel I cont ( x, y ) as the ratio between pixel’s locally sensed light intensity I ph ( x, y ) and a spe-
1. This chapter of the thesis is based on a previous reported paper ([23]). I would like to thank the anonymous reviewers for improving the quality of the paper.
105
Chapter 4: The AER Spatial Contrast Vision Sensor
cially weighted average of its surrounding neighborhood computed with some kind of diffusive
network
I ph ( x, y )
I cont ( x, y ) = I ref ---------------------I avg ( x, y )
(4.1)
where I ref is a global scaling current. Since this is always positive, let us call it “unipolar”
contrast computation, with contrast being computed as the ratio between two photo currents. This
yielded circuits where no subtraction operation was required. This was crucial to maintain
mismatch (and precision) at reasonable levels. Note that for computing I avg and I cont circuits
have to handle directly photo currents, which can be as low as pico-amperes or less. Performing a
simple mirroring operation introduces mismatches with errors in the order of 100% [14]. This can
be overcome by increasing transistor area, but then leakage currents may become comparable to
the available photo currents. Consequently, while handling photo currents, it is desirable to keep
complexity at a minimum. Therefore, from a circuit point of view, the way of computing contrast
as in eq. (4.1) was very convenient. However, this presents an important drawback: when there is
no contrast ( I avg = I ph ) then I cont ≠ 0 . In an AER circuit this means that a pixel sensing no
contrast will be sending out information (events) and consuming communication bandwidth on
the AER channels. This is contrary to the advantages of AER (where it is expected that only
information relevant events will be transmitted) and contrary to the advantages of computing
contrast at the focal plane (so that only contrast relevant pixels need to send information). In
another work [14], although spatial contrast was computed by eq. (4.1) in the retina, a postprocessing with AER (convolution) modules was added to effectively compute the Weber
Contrast2 as the signed quantity
I ph ( x, y )
I cont ( x, y ) = I ref  ---------------------- – 1
 I avg ( x, y ) 
(4.2)
This reduced significantly the data flow (from about 400keps3 to about 10keps), but also at the
expense of reducing the speed response of a pixel by a factor of about 10.
In this chapter, we present a new spatial contrast retina design [23], where the contrast computation follows eq. (4.2). The design is based on the original contrast computation circuit by
Boahen [45], which has been improved to overcome its inherent limitations on mismatch, ambient
light dependence, and critical controllability. It has also new functionalities. We have signed output. Furthermore, and adjustable thresholding mechanism has been included, such that pixels
remain silent until they sense an absolute contrast above the adjustable threshold. The retina also
includes an optional global reset mechanism for operation in ambient-light-independent Time-toFirst-Spike Contrast Computation Mode. A 32 x 32 pixel test prototype was fabricated in 0.35 µm
CMOS technology.
2. Weber Contrast is defined as WC = (I-Iavg)/Iavg for a pixel photo current with respect to its neighbourhood average photo current, or as WC = (I1-I2)/(I1+I2) between two adjacent pixels or regions. Both
expressions are equivalent by making I = I1 and Iavg = (I1+I2)/2.
3. keps stands for “kilo events per second”.
106
Chapter 4: The AER Spatial Contrast Vision Sensor
Iref
Translinear Multiplier
Iph(x,y)
to (x,y+1)
to (x−1,y)
M5
I avg (x,y)
Diffusive
Network
to (x,y−1)
M1
V slc
Icont (x,y)
to (x+1,y)
I photo (x,y)
M2
I bias
I ph (x,y)
M3
Fig 4.1: Block diagram of pixel in prior unipolar contrast retina.
4.2 Prior AER Mismatch-Calibrated Unipolar Spatial Contrast AER Retina
The AER contrast vision sensor is based on a previous design [14] presented by Costas et al.
in 2007. Figure 4.1 shows the basic schematic of the contrast computation circuit used in the previous unipolar spatial contrast retina. A p+/nwell photo diode senses current I ph ( x, y ) that is
replicated twice using a sub-pico-ampere current mirror [38]. The first replica is used in a cascoded diffusive network [46], which implements the discrete approximation of the 2D Laplacian
equation
2
2

∂
∂ 
I ph ( x, y ) =  1 – λ x 2 – λ y 2  I avg ( x, y )

∂x
∂y 
(4.3)
This equation provides a good spatial average of I ph over neighboring pixels, such that closer
pixels contribute more to this average than distant pixels. The second replica of the photo current
is fed together with I avg ( x, y ) to a translinear circuit computing the ratio between both, scaled by
107
Chapter 4: The AER Spatial Contrast Vision Sensor
AER
AER
Channel−1
Retina
Merger
Constant
Image
AER
Channel−3
Convolution
Chip
AER
Channel−4
AER
Channel−2
Fig 4.2: Setup used to convert the unsigned AER retina output with DC level to a signed
AER stream with no DC level.
reference current I ref . The resulting current I cont ( x, y ) is thus proportional to a unipolar contrast (as in eq. (4.2)) and is fed to an integrate-and-fire neuron generating periodic spikes with a
frequency proportional to I cont ( x, y ) . Scaling current I ref is made locally trimmable for each
pixel in order to compensate for all mismatch contributions from the photo diode, current mirror,
diffusive network, translinear circuit, and integrate-and-fire neuron. As a result, inter-pixel mismatch contrast computation could be reduced from about σ ≈ 60% to σ ≈ 6% using 5-bit
pixel registers to control I ref ( x, y ) . Pixel complexity was kept relatively simple (104 transistors
+ 1 capacitor) thanks to the unipolar nature of the contrast computation, and the whole pixel could
be fit into an area of 58µm × 56µm in a 0.35µm CMOS process. The main drawback is that
pixels with no contrast would generate output events at a constant rate proportional to I ref . To
overcome this, the 4-AER-module system shown in Figure 4.2 was assembled to compute effectively a bipolar contrast as in eq. (4.2). A uniform image AER flow with negative sign bit was
merged with the retina AER flow and fed to an AER convolution chip [27] configured to operate
as an array of signed (bipolar) integrators. As a result, the background DC component in eq. (4.1)
was removed, yielding a computation equivalent to that in eq. (4.2). However, as a backside
effect, the effective firing rate of a pixel at the output channel was reduced by a factor of 8, thus
diminishing its speed response. In the design presented in this chapter, this is solved by performing all the bipolar contrast computation at the sensor chip using an improved version of Boahen’s
original biharmonic contrast computation circuit.
4.3 Boahen Spatial Contrast Pixel
In the spatial contrast sensor presented in this thesis all these drawbacks are solved by performing all the signed-spatial-contrast computation at the sensor chip using an improved version
of Boahen’s pixel circuit [46]. The continuous approximation of Boahen’s pixel circuit, shown in
Figure 4.3 solves approximately the following equations [46]
108
Chapter 4: The AER Spatial Contrast Vision Sensor
Ih
Mh
Vhh
Mc
Vcc
C
Ma
Ih
Iu
Mh
H
Mb Vhh
Mc
Iph
Ma
C
Ic
Vcc
Iu
H
Mb
Iph
Ic
Fig 4.3: Boahen original contrast computation circuit.
2
I h ( x, y ) = I ph ( x, y ) + a∇ I c ( x, y )
2
I c ( x, y ) = I u – b∇ I h ( x, y )
(4.4)
Solving for I h results in the biharmonic equation used in computer vision to find an optimally smooth interpolating function of the stimulus I ph . Consequently, the output I c ( x, y ) is
the second order spatial derivative of the interpolation I h according to the bottom eq. (4.4). Since
the interpolation is a spatially integrated version of the stimulus, I c can be interpreted as a version of a first order derivative of the stimulus, therefore, spatial contrast. This can also be understood with the help of Figure 4.4. The top trace shows a step stimulus I ph and its spatial average
( I avg or I h ). The center trace shows the contrast computation as I avg ⁄ I ph , and the bottom
trace shows the contrast computation as the second order spatial derivative of I h . Both are equivalent. According to the bottom eq. (4.4), I c includes a DC term I u .
Obtaining an expression for each pixel output from equations (4.4) is not trivial and intuitive. Each pixel output current depends on all the pixels of the retina. An approximation given by
Boahen [46] for each pixel output from equations (4.4) is
I ph ( x, y )
I c ( x, y ) = I ref --------------------------------------〈 I ph〉 + I ph ( x, y )
(4.5)
where 〈 I ph〉 is the average photo current in the pixel’s surrounding neighborhood. Obviously,
closer neighborgs will have higher influence on the output and far pixels will not have a noticeable effect. By adjusting Icc, neighbors influence can be controlled.
109
Chapter 4: The AER Spatial Contrast Vision Sensor
1
I
ph
Iavg or Ih
0.75
0.5
2
Iavg/Iph
1.5
1
0.5
1
dIh/dx
0
2
2
d Ih/dx
−1
Fig 4.4: Interpretation of spatial contrast computations.
The original circuit implementation of this model suffered from a series of drawbacks. First,
mismatch was comparable to output signal. Second, output signal would degrade for the same
contrast stimulus when changing lighting conditions. Third, bias voltages V cc and V hh in Figure
4.3 had very narrow and critical tuning range. All three drawbacks have been improved with the
present implementation.
4.4 Improved Signal-Spatial-Contrast Pixel
Figure 4.5 provides an overall block diagram, indicating the signals interchanged between
blocks. The pixel contains three main parts: (1) the photo sensing and contrast computation part,
(2) the calibration part, which provides the ambient light independent contrast current I cont ; (3)
the integrate-and-fire part, which includes refractory circuitry, thresholding, and TFS mode; (4)
and the pixel AER communication circuitry that sends out events to the periphery. Let us now
describe each one.
A.
Compact Calibration Circuit
As we explained previously, one of the main drawbacks of prior AER contrast retinae was
high mismatch. To overcome that problem, we reduce pixel mismatch by introducing calibration.
One dominant source of mismatch is the DC component I u in eq. (4.4). Since this current is set
110
Chapter 4: The AER Spatial Contrast Vision Sensor
Vrfr
Mh
Icont
Mc
i&f
Ical
Iph
calibration
Photo sensing and
contrast computation
to event read out
_
col +
col
TFS
pulse+
ev_rst +
_
pulse
_
ev_rst
+
−
Vth
Vth
integrate−and−fire
with thresholding
TFS reset and
refractory time
+ event block
Rqst
_
Ack
event block
to row
arbiter
AER pixel
communication
Fig 4.5: Overall block diagram of the contrast pixel.
VDD
D
bN−1
S N−1
G
VG
wcal
b1
S1
b0
VD
VS
S0
S
Fig 4.6: Digitally controlled length MOS used for calibration and its symbol.
constant, independent of lighting conditions, we can directly subtract it with a trimmable current
source. The output current will thus be directly the bipolar contrast current we were looking for.
To implement the trimmable current source, we follow the explained compact circuit based on
series transistors association [13] previously explained. Figure 4.6 shows the basic principle
behind this circuit. Each switched MOS operates as a segment of an effective longer MOS whose
length is controlled digitally by switching individual segments from ohmic to saturation, and vice
versa. The key consists in making each segment to contribute approximately as a power of 2 to the
total length. The digital control word w cal = { b N – 1 …b 1 b 0 } sets the state of the switches. As
a result, the effective length is digitally controlled as in a digital-to-analog conversion. On the
right of Figure 4.6 we show the symbol of a digi-MOS (digitally-controlled MOS) which we use
to represent the circuit on the left.
To calibrate the circuit, we use the second approach described in chapter 3, [13]. This
scheme has higher precision at the calibration point but calibration degrades faster when bias cur-
111
Chapter 4: The AER Spatial Contrast Vision Sensor
M3 I3n
M1
I2
I1
M2
I3
Ical
M4
Mq
Mp
wcal
I’u
Vbu
Fig 4.7: Translinear tuning circuit used to calibrate the contrast retina.
rents are scaled. In our particular case, we were interesting in achieving the highest precision as
the bias current Iu value will not be scaled significantly after calibration.
Figure 4.7 shows the circuitry used to subtract the DC component I u of the contrast current.
Transistors to the left of the dashed line are shared by all pixels and are located at the chip periphery, while those to the right are replicated for each pixel. Current I u' sets the subtracting DC level
(while also introducing mismatch), while I cal is adjusted to a tuning range covering the interpixel
mismatch. Transistors M 1 – 4 form a translinear loop [46], thus I cal = I 1 I 2 ⁄ I 3n . And I 3n is a
mirrored version of I 3 by transistors M p and M q . Transistor M q is digi-MOS of Figure 4.6.
Consequently, I cal is proportional to the pixel calibration word w cal ( x, y ) , which is stored on
in-pixel static RAM latches loaded at start-up. Note that current I cal could have been generated
directly by current mirror M p - M q . However, in this case, if one wants to scale { I u , I cal , I u' }
globally (to adjust the retina output frequency range) the circuit would change the current through
the calibration branch containing M q . On the contrary, with the circuit in Figure 4.7 one can
scale { I u , I cal , I u' }while keeping the calibration branch current I 3n (and I 3 ) constant, and
scale through peripheral currents I 1 and/or I 2 . This way, calibration degrades less when tuning the
output frequency range.
B.
Photo Sensing and Contrast Computation
Figure 4.8 shows how Boahen’s contrast computation circuit has been modified to include a
current biasing scheme for controlling the original voltages V cc and V hh in Figure 4.3. This
way, gate voltages V cc and V hh tend to follow voltage excursions at nodes ‘C’ and ‘H’.
The first advantage of this is that biasing will adapt to ambient light conditions. For example, if all photodiode currents are scaled up/down by the same factor, the voltage at all nodes ‘H’
will follow it logarithmically. Since I u is constant, the voltage at node ‘C’ will thus also follow
the same shift. Since bias currents I hh and I cc are kept constant, the gate voltages of transistors
112
Chapter 4: The AER Spatial Contrast Vision Sensor
Ih
Mh
Ihh
Mc
Ma
H
Mb I
hh
C
Iph
Icc
I’u
Ih
Iu
Mh
Mc
Ic
Icc
Ma
C
Iu
H
Mb
Iph
Ic
Icont
Vb
wcal
Ical
Fig 4.8: Detail of photo sensing and contrast computation circuit.
M h and M c will thus follow also this same global voltage shift, adapting themselves to the global
light change.
The second advantage of this current biasing scheme is that it attenuates mismatch. After
doing careful mismatch analysis and identifying the main sources of mismatch for this circuit, one
can find out that transistor M a and current I u are the dominant sources of mismatch. This can be
understood as follows. Mismatch in I u goes directly into the DC offset of I c , which will be calibrated by I cal . Mismatch of M b is less critical because its inter-pixel gate voltage (node ‘C’)
variability affects the bottom diffusive grid and the computation of the average current I h . Thus
its variability impact is attenuated by the average computation. However, M a mismatch ( V gs
variation of M a ) changes directly the source voltage of M b , affecting directly the gain of contrast output (coefficient ‘b’ in eq. (4.4)), whose effect is not directly calibrated by I cal . Consequently, M a needs to be sized to minimize mismatch. The effect of I u will be compensated by
calibration, and the effect of M a will be attenuated by the current biasing scheme. Note that mismatch in all M a transistors will introduce random voltage variations at nodes ‘H’ and ‘C’. These
variations will be transformed into random lateral currents through transistors M h and M c . The
random currents through M h will be collected by output current I c and can be compensated by
calibration. However, random currents through M c transistors operate as if they were generated
by the photo diodes. Thanks to the current biasing scheme, an increase in ‘C’ will increase the
113
Chapter 4: The AER Spatial Contrast Vision Sensor
gate voltage of the new bottom NMOS transistor, increasing its source voltage, thus increasing the
gate voltage of M c , which will reduce the lateral random current. A similar effect will be happening for transistors M h .
Finally, the third advantage is a more robust means for biasing the lateral transistors. In the
original scheme, voltages V cc and V hh suffered from a very narrow and critical tuning range
(about 100mV or less). Now, bias currents I cc and I hh can be tuned over several decades, while
still perceiving their effect.
C.
Integrate-and-Fire
Figure 4.9(a) shows the integrate-and-fire block. Input contrast current I cont is integrated on
capacitor C int . Two comparators detect whether the capacitor voltage V cap reaches an upper
( V high ) or lower ( V low ) threshold, triggering the generation of a positive (pulse+) or negative
(pulse-) event, respectively. To accelerate the comparisons, both comparators activate a positive
feedback loop (from V cap to V dd04 for a positive event, or from V cap to V gn04 for a negative
event).
After event generation, capacitor C int is reset to the central voltage V ref . This is done by
the reset circuit shown in Figure 4.9(b). This reset mechanism includes a refractory timing circuit
that inhibits the pixel from generating subsequent events before refractory capacitor C rfr has
been discharged by the DC current source MOS controlled by V rfr . The reset circuit also
includes the global TFS (Time-to-First-Spike) mode reset signal, which resets all pixel capacitors
C int simultaneously. Note that this signal inhibits the positive feedback loops in Figure 4.9(a).
This allows resetting quickly those pixels generating an event when TFS becomes active.
Figure 4.9(c) shows the minimum contrast thresholding circuit. A comparator detects
whether capacitor voltage is above or below V ref and turns on either a positive ( I low ) or negative
( I high ) threshold current, which I cont needs to exceed for producing an event. Figure 4.10 shows
the resulting relationship between integrate-and-fire circuit output frequency f out and the input
+
_
signed contrast current I cont while bias voltages V th and V th are set to generate threshold currents I high and I low, respectively. Naturally, threshold transistors would also introduce mismatch.
Consequently, they were layed out with a large area of 2 ⁄ 20 µm .
Figure 4.9(d) shows the two-stage comparators used in Figure 4.9(d). At stand by they are
biased at low current through V b1 and V b2 . However, during event generation its bias current is
increased. This increase starts when signals pulse starts to depart from its resting voltage and
stops after the pixel event reset signal ev_rst returns to its resting level. The comparator within the
thresholding circuit in Figure 4.9(d) does not have this feature, since this comparator only needs
to detect whether the so far accumulated contrast for the pixel is positive or negative, which is a
slow process compared to the event generation timings.
114
Chapter 4: The AER Spatial Contrast Vision Sensor
(a)
Vdd04
TFS
V ref
Icont
Vrfr
ev_rst +
pulse+
Vhigh
pulse+
ev_rst +
TFS
Vcap
Reset
Rfrct
Cint
ev_rst +
Thres−
hold
ev_rst
_
+
V th V th
_
ev_rst
_
ev_rst _
pulse
_
_
pulse
TFS
Vgn04
(b)
Vcap
Vlow
(c) Ihigh
V rfr
Crfr
_
Vth
Vref
ev_rst +
_
ev_rst
TFS
Vref
Vcap
I
low
+
Vth
(d) ev_rst
pulse
Vb
1
Vb2
_
Vi
+
Vi
Vo
Fig 4.9: (a) Integrate and fire neuron. (b) Reset and refractory circuit. (c) Thresholding
circuit. (d) Detail of comparators.
115
Chapter 4: The AER Spatial Contrast Vision Sensor
fout
I high
I low
Icont
Fig 4.10: Effect of contrast thresholding on the relationship between pixel output frequency and contrast current.
col
Rqst
pulse
Ack
ev_rst
Fig 4.11: AER pixel communication blocks.
D.
Communication Block Circuit
Finally, the AER pixel communication part in Figure 4.5 contains two identical “event
block” circuits, which are shown in Figure 4.11 These are standard AER pixel communication circuits taken from Boahen’s row parallel event read-out technique [47]. When generating signed
116
Chapter 4: The AER Spatial Contrast Vision Sensor
Table 4.1: Chip Specifications.
technology
power supply
chip size
array size
pixel size
fill factor
photodiode quantum
efficiency
pixel complexity
current consumption
CMOS 0.35µm 4M 2P
3.3V
2.5 x 2.6 mm2
32 x 32
80 x 80 µm2
2.0%
0.34 @ 450nm
131 transistors + 2 caps
65µA @ 10keps
events, each pixel needs to provide two column event signals col+ and col-. This concept was
already implemented and tested in prior designs [27] that required signed events.
4.5 Experimental Results
A 32 x 32 pixel test prototype AER signed spatial contrast retina chip was designed and fabricated in a double poly 4-metal 0.35µm CMOS process with a power supply of V DD = 3,3V .
Table 4.1 summarizes the chip specifications. Figure 4.12 shows a micro photograph of the die, of
size 2.5 x 2.6 mm2. The whole chip, except the pad ring, is covered with the top metal layer leaving openings for the photo diode sensors. Figure 4.12 also shows the layout of a single pixel highlighting its components. Each pixel layout is a symmetrical reflection of its neighboring pixels.
This way noisy digital lines are shared among neighbors, as well as power supplies, and noise
sensitive bias lines among other capabilities [52]. At the same time, noise sensitive lines are separated from noisy ones. Pixel area is 80 × 80µm 2 , including routing. The pixel was made up of
131 transistors and 2 capacitors (the capacitance of the integrate-and-fire circuit and the capacitance of the reset and the refractory circuit).
Figure 4.13 shows the experimental setup to characterize the temporal contrast sensor. The
chip was mounted on the Caviar Board (a specific PCB designed to test generic AER devices).
This board was connected to a USB port. The digital words to program the bias currents of each Ipot [37] were sent through this port. The outputs of the Caviar Board were sent to an AER bus.
This bus was connected to a Mapper. The Mapper was a programmable device that was used to
filter or remove the activity of some pixels for some specific measurements (latency characterization), without affecting the activity of the rest of the pixels. The output bus of the Mapper was
connected the board USB2AER, [52], which sends the AER data from the bus to one of the USB
computer ports. This information was processed by jAER [49] that allows to see real time data.
The USB2AER board was specially useful for real-time monitoring. The input bus to this board
was replicated and sent to a Datalogger. This board has an internal memory that can store the data
generated by 524.000 events and was used to save efficiently on the PC the AER information
transmitted from the AER bus.
117
Chapter 4: The AER Spatial Contrast Vision Sensor
2.6mm
32x32 pixel array
Row Arbiter
2.5mm
Bias Generator
AER-out
Configuration Registers
digital
comm.
i&f
80µm
thresholding
photodiode
calibration
contrast
computation
Fig 4.12: Microphotograph of 2.5mm x 2.6mm die, and zoom out of 80µm x 80µm pixel
(layout) indicating the location of its components.
118
Chapter 4: The AER Spatial Contrast Vision Sensor
usb
Mapper
USB2AER
AER BUS
usb
AER BUS
Datalogger
Ipots
Caviar Board
usb
Datalogger
Lens Mount
Mapper
Caviar Board
USB2AER
Fig 4.13: Experimental setup and photo of the retina and main boards used to test the
imager.
119
Chapter 4: The AER Spatial Contrast Vision Sensor
4.5.1 Pixel Frequency Range
In order to control the pixel output frequency, one of the corner pixels had its integrating
capacitor node connected to a low-input-capacitance analog buffer for monitoring purposes. Pixel
integrating capacitors have a capacitance of about C int ≈ 118fF (obtained from the layout
extractor), while the corner pixel with monitoring buffer has a total capacitance of about
C mntr ≈ 196fF (estimated from layout extraction and simulation). Figure 4.14 shows recorded
waveforms (for positive and negative currents) for this capacitor when turning off horizontal
interactions among neighboring pixels (by turning off transistors M h and M c in Figure 4.8),
and for a typical value of I u ≈ 100pA . By changing I u (with I u' = I cal = 0 ) or I u' (while
I u = I cal = 0 ), pixel oscillation frequency could be tuned between 1.2Hz and 5KHz. For the
maximum frequency the arbitrating periphery inserts varying delays. This is because all pixels are
also firing with maximum frequency (even higher than the pixel we are observing which has
slightly higher integrating capacitance) and are collapsing the arbiter. Consequently, in a practical
situation where only a small percentage of the pixels would fire with maximum frequency, they
would be able to fire with a higher than 5KHz maximum frequency.
4.5.2 Calibration
Although the retina can work without calibration, in order to use have bipolar output and
reduce mismatch, the first requirement is to calibrate it. For this, the retina was exposed to a uniform stimulus, while biased for normal operation conditions but without thresholding. In our case,
normal operation conditions are I u = 150pA , V ref = 1,65V , V high = 2,8V , V low = 0,45V ,
I hh = 10pA , I cc = 5pA . Also, before calibration, we set I cal = I u' = 0 . Under these conditions, retina output events are recorded, from which one can obtain the firing frequency of each
pixel. Next, we set current I u' = 80pA so that the pixel with minimum frequency has a frequency close to zero (or slightly negative). Under these conditions the resulting histogram of pixel
frequencies distributions is shown in Figure 4.15.(a) After this, the calibration circuit biases
( I 1, I 2, I 3 in Figure 4.7) were set for optimum coverage of this distribution, and for each pixel the
optimum calibration word w cal ( x, y ) was found. This is computed off-line by optimally combining biases { I 1, I 2, I 3 } and calibration words w cal ( x, y ) , following a pre-established optimization
criterion. In our case, we allowed for a few outliers in order to minimize the residual standard
deviation. One could also target to minimize the spread among the most extreme pixels at the
expense of a higher standard deviation. After this process, the histogram of calibrated pixel frequencies obtained is shown in Figure 4.15(b). The residual inter-pixel standard deviation is 26Hz.
As we will explain later, maximum contrast frequency for these biases is ± 4400Hz . Consequently, post-calibration residual mismatch is 0.6%. Figure 4.16 shows the calibration ladders
obtained for the 1024 pixels that compounds the retina. We have also plotted with red asterisks the
optimum output frequencies after calibration versus the optimum calibration word for each pixel.
Finally, in Figure 4.17 we have plotted the output frequencies before calibration (with a bipolar
output subtracting the current Iu) and after calibration. Output frequencies were calculated counting the pixels spikes within a time interval. Datalogger [48] was used to save the number of
spikes generated for each pixel during the interval.
120
Chapter 4: The AER Spatial Contrast Vision Sensor
Positive events (Minimum Output Frequency)
Voltage (V)
3
2.5
2
1.5
0
2
4
6
8
10
Voltage (V)
Negative events (Minimum Output Frequency)
1.8
1.3
0.8
0.3
0
2
4
6
8
10
Time (s)
Positive Events (Typical Output Frequency)
Voltage (V)
3
2.5
2
1.5
0
5
10
15
20
25
30
Voltage (V)
Negative events (Typical Output Frequency)
1.8
1.3
0.8
0.3
0
5
10
15
Time (ms)
20
25
30
Positive events (Maximum Output Frequency)
Voltage (V)
3
2.5
2
1.5
0
0.5
1
1.5
2
Voltage (V)
Negative events (Maximum Output Frequency)
1.8
1.3
0.8
0.3
0
0.5
1
Time (ms)
1.5
2
Fig 4.14: Recorded waveforms at the integrating capacitor. Minimal oscillation frequency
(1.2 Hz), Typical oscillation frequency (466 Hz), and maximum oscillation frequency (5
KhZ).
121
Chapter 4: The AER Spatial Contrast Vision Sensor
80
60
(a)
Ical=0
40
20
0
−400
−200
0
200
400
600
400
Optimized Ical
(b)
200
0
−400
−200
0
200
Frequency (Hz)
400
600
Fig 4.15: Histograms of retina pixels frequencies distribution before and after calibration
at the same illumination level.
Figure 4.18 shows how the standard deviation of the post-calibration residual mismatch
changes with illumination level. The figures shows three superimposed graphs. Each correspond
to performing calibration at different illumination levels {50, 5, 1, and 0.25 Klux}.The worst case
situation corresponds to calibrating at about 1k-lux and using the retina at very high light conditions, resulting in a standard deviation of almost 140Hz (3%). On the other hand, the optimum situation corresponds to calibrating at 15k-lux, which results in a standard deviation of less than
80Hz (1.8%) over the entire 5 decade range.
The calibration process is all done off-line. However, it is conceivable to implement it fully
on-chip (through, for example, a described state machine), since it only requires to expose the
chip to uniform illumination (one can simply remove the optics), compare the pixel frequencies
(for which not even a precise clock reference is required), and compute an optimum set of calibration weights.
4.5.3 Contrast Step Response
Figure 4.19 illustrates the retina response to a luminance step of different contrast levels,
while thresholding is turned off. Input stimulus is printed paper, providing a static image with a
half dark and a half grey side. The half gray side intensity is adjusted between 100% (white) and
30% (darkest gray). Table 4.2 indicates the relationship of the luminance steps, with the ratio of
photo currents between the gray and black parts, and the resulting Weber Contrast, defined as
122
Chapter 4: The AER Spatial Contrast Vision Sensor
1000
800
600
Frequency (Hz)
400
200
0
−200
−400
−600
−800
−1000
0
5
10
15
Wcal
20
25
30
Fig 4.16: Calibration ladders and optimum calibrate words for all the pixels of the retina.
I light – I dark
WC = ----------------------------I light + I dark
(4.6)
The left column in Figure 4.19 shows this input stimulus image. The center column in Figure
4.19 shows the retina output response before calibration, while the right column shows the retina
response after calibration. Central gray level is zero pixel frequency. Brighter pixels are firing
positively signed events, while darker pixels are firing negatively signed events. Absolute maxi-
Table 4.2: Measured luminance steps and Weber contrast.
luminance step
100%
to 0%
70%
to 0%
50%
to 0%
30%
to 0%
10%
to 0%
0% to
0%
Ilight/Idark
9
6
3.6
2.4
1.5
1
Weber Contrast
(WC)
0.80
0.72
0.56
0.41
0.20
0
123
Chapter 4: The AER Spatial Contrast Vision Sensor
800
Before calibration
After calibration
600
Frequency (Hz)
400
200
0
−200
−400
−600
−800
0
200
400
600
Pixel number
800
1000
Fig 4.17: Output frequencies before and after calibration.
mum pixel frequency was 250Hz. Biasing conditions in Figure 4.19 were I u = 150pA ,
I u' = 150pA ,
V high = 2.9V , V low = 0.4V , and V ref = 1,65V .
4.5.4 Contrast Sensitivity
An important characterization for a spatial contrast retina is its contrast sensitivity what is
the output event rate for a given input contrast stimulus. We have characterized spatial contrast
sensitivity for the positive event branch and the negative event branch (see Figure 4.5) separately,
since they have separate circuitry. Usually, under normal operation, the retina will be biased to
have the same sensitivity for positive and negative events. However, there might be situations
where one would prefer to set different contrast sensitivities for positive and negative events, and
this retina offers this possibility. To characterize pixel contrast sensitivity, a gray level step stimulus (as shown in Figure 4.19) of different contrast values, was used. Pixels frequencies of the two
columns with the highest activity (the ones just on the left and right of the stimulus center) were
recorded. This process was repeated for different bias values for V high and V low , with
124
Chapter 4: The AER Spatial Contrast Vision Sensor
140
120
50 Klux
15 Klux
5 Klux
1 Klux
250 lux
Deviation (Hz)
100
80
60
40
20 0
10
1
10
2
3
10
10
Illuminance (lux)
4
10
5
10
Fig 4.18: Effect of ambient illumination on post-calibration residual mismatch standard
deviation. Three curves are shown, each for calibration at the given illumination level.
V ref = 1.65V . The results are shown in Figure 4.20(a). The measured maximum contrast sensitivity was 4400Hz/WC (Hz per Weber Contrast) for V high – V ref = V ref – V low = 0.15V .
Error bars indicate inter-pixel variability.
To show the sensitivity dependence with illumination, the maximum output frequency for a
Weber Contrast of WC = 0.8 was measured (for both signs of contrast) with different illumination
levels. As shown in Figure 4.20(b), sensitivity degrades slightly when illumination decreases.
Sensitivity remains almost constant over the first two decades, and approximately doubles over
the second two decades.
4.5.5 Contrast Thresholding
In Figure 4.21, the typical pixel output when the visual field is swept with a grey level bar
stimulus of WC = 0.8 is shown. The x-axis indicates bar position in row number units. The pixel
output spike frequency reaches the maximum value when the stimulus is at the pixel´s row. This
value depends on the width of the sweeping bar. Several outputs using different bar widths have
been plotted for the same pixel. The bar width is expressed in projected pixel units. The maximum
frequency is proportional to the stimulus width. In both cases, the following voltages were used:
125
Chapter 4: The AER Spatial Contrast Vision Sensor
before
calibration
after
calibration
0%-30%
WC=0.41
0%-50%
WC=0.56
0%-70%
WC=0.72
0%-100%
WC=0.8
input
Fig 4.19: Retina response to a luminance step changing Weber Contrast. Left column is
input stimulus. Centre column is output response before calibration, and right column is
output response after calibration.
V high = 2.9V ,
V low = 1.4V
and
V ref = 1.65V .
With
these
settings,
V high – V ref > V ref – V low , so negative events were enhanced.
It is also possible to fully inhibit positive or negative events by setting either I high or I low
(see Figure 4.9(c)) to sufficiently large values. Asymmetrical thresholds ( I low ≠ I high ) can also
be used. Therefore, positive and negative events can be inhibited independently. In Figure 4.22,
the effect of thresholding is shown. First, the visual field was swept with a 100% contrast bar for
different thresholds. Figure 4.22(a) shows the output frequency for pixel (17,11) when setting
symmetric thresholds. Figure 4.22(b) shows the same pixel results but when setting only threshold values to inhibit positive events. The negative output frequency remains constant.
The main advantage of thresholding is to remove the residual mismatch after calibration.
Pixels usually spike with a low residual output frequency after calibration. Positive and negative
thresholds can be set to remove these undesirable outputs after calibration. Figure 4.22(c-e) show
126
Chapter 4: The AER Spatial Contrast Vision Sensor
Contrast Sensitivity
Frequency (Hz)
4000
2000
Threshold=0.15V
Threshold=0.2V
Threshold=0.5V
Threshold=0.8V
0
−2000
−4000
−0.1
0
0.1
0.2
0.3
0.4
0.5
Weber Contrast
0.6
0.7
0.8
0.9
Dependence of the Sensitivity with Illumination, WC=0.8
Sensitivity (Hz/WC)
5000
Max
Min
4000
3000
2000
1000 1
10
2
10
3
10
Illumination (lux)
4
10
5
10
Fig 4.20: Contrast sensitivity measurements. A stimulus step (as in Figure 4.19) was applied
and max and min frequencies were recorded. Top panel shows mas and min frequencies for
different stimulus step contrasts and different threshold values. Bottom panel shows how
the maximum and minimum frequencies depends on illumination (WC=0.8).
127
Chapter 4: The AER Spatial Contrast Vision Sensor
200
0
Frequency (Hz)
−200
−400
BarSize=1.6 pixels
BarSize=1.9 pixels
BarSize=2.2 pixels
BarSize=2.5 pixels
BarSize=3.2 pixels
BarSize=4.0 pixels
−600
−800
−1000
−1200
−1400
0
5
10
15
20
Stimulus Position
25
30
Fig 4.21: Typical pixel’s output when the retina is stimulated with a 100% contrast bar of
different widths.
some snapshots captured with the contrast retina. Central gray color indicates zero output (no contrast). Positive events range from this gray to black and negative events range from this gray to
white. The three snapshots were taken for different values of the positive and negative thresholds.
For the three cases, I u = 150pA . In Figure 4.22(c) a positive threshold current of 1nA was set
to inhibit positive events completely after calibration. I low was 150pA. In Figure 4.22(d) a symmetric threshold of 80pA was set after calibration. In Figure 4.22(e) the retina output without neither calibration nor thresholding is shown. Above each snapshot the sum of all pixels’ frequencies
f total is indicated. We can see, by comparing (d) and (e), that calibration reduces event flow
(communication bandwidth) while enhancing contrast gain.
4.5.6 Latency Characterization
Sensor latency is the time it takes from the occurrence of the illumination change at the
photo diode to the output of an address event. Sensors with quick answers to stimuli are suitable
to be used as part of a feedback controller. Latency depends on the pixel, the asynchronous bus
arbitrer and the periphery.
To characterize the retina latency we proceeded as follows. We stimulated a LED with a step
signal to turn it ON, focused it over a central region of the sensor array, and recorded the time
delay between the step signal and the first event Rqst coming out of the pixel under test from that
region. The Mapper was programed to allow spike only the pixel under test of the stimulated
128
Chapter 4: The AER Spatial Contrast Vision Sensor
Positive and Negative Threshold
Frequency
400
0
−400
Thresh=0pA
Thresh=120pA
Thresh=200pA
Thresh=300pA
Thresh=400pA
−800
−1200
(a)
−1600
0
5
10
15
20
25
30
Only Positive Threshold
Frequency
400
0
−400
Thresh=0pA
Thresh=120pA
Thresh=200pA
Thresh=300pA
Thresh=400pA
−800
−1200
(b)
−1600
0
(c) ftotal=1.3 x 105 Hz
5
10
15
20
Stimulus Position
(c) ftotal=1.8 x 105 Hz
25
30
(e) ftotal=3.3 x 105 Hz
Fig 4.22: Effect of thresholding. (a) Bar is swept for different symmetric thresholds. (b) No
threshold for negative events, and positive event thresholds are changed. (c) Events captured for calibrated retina when all positive events are inhibited by setting a high positive
threshold. (d) Events captured for calibrated retina with symmetric threshold. (e) Events
captured for uncalibrated retina.
region. An oscilloscope working in single mode was used to measure the latency. The measurements were repeated by inserting different neutral density filters to attenuate light intensity from
about 50k-lux down to 2 lux. The resulting latencies are shown in Figure 4.23. The measurement
was repeated by focusing the LED over different regions of the pixel array. The bars in Figure
4.23 show the spread obtained when repeating measurements. As can be seen, latency changes
129
Chapter 4: The AER Spatial Contrast Vision Sensor
−1
10
−2
Latency (s)
10
−3
10
−4
10
−5
10
0
10
1
10
2
3
10
10
Illumination (lux)
4
10
5
10
Fig 4.23: Latency measurements under changing illumination conditions.
from about 10ms down to about 0.1ms when illumination varies over almost 5 decades. This
means that latency is dominated by the photo sensing circuits. However, latency does not scale
proportionally to light, and consequently this retina does not suffer from the severe Light-to-Time
restriction listed in Table 1.1.
4.5.7 Natural Scenes
Although the retina resolution is rather low (32 x 32 pixels) for observing natural scenes,
Figure 4.24 shows some captured images when observing natural elements, which give a first
order feeling of how an up-scaled retina would respond under a natural scene. jAER [49] was
used to convert AER data into a sequence of frames for real-time visualization and monitorization. We used a lens objective of 16mm.
4.5.8 TFS Output Mode
As mentioned on page 112, the integrate-and-fire circuit of the retina pixel can be configured to operate in TFS mode. In this mode, the refractory period of the retina has to be set to its
largest possible value (by connecting voltage V rfr to V dd ) to guarantee that each pixel will fire at
the most one single event. Then a periodic reset pulse has to be provided for global signal TFS .
130
Chapter 4: The AER Spatial Contrast Vision Sensor
Fig 4.24: Natural elements. From left to right: screw, paper clip, eye and child face. We
used a lens objective of 16mm.
This can be done in several ways. One trivial option is to reset at a fixed preset frequency. However, another more efficient option is by counting the output events. Since output events are coming out in decreasing order of pixel contrast, high contrast pixels (either positive or negative)
come out first. These are the pixels carrying more relevant information, for example, for a recognition application. Consequently, one could add a simple counter at the Rqst line and have it generating a reset pulse for TFS after each M events. This way, a dynamic “frame time” T frame would
131
Chapter 4: The AER Spatial Contrast Vision Sensor
Fig 4.25: Paper clip snapshots in TFS mode for different number of captured events, M.
TM
TFS
Tfirst
Rqst
1
2
Tframe
M−1 M
Fig 4.26: Time line of the Global Reset and the Request signal.
be produced which self adjusts to the contrast level of the scene, independent of ambient light.
High contrast scenes would self-tune to faster frames, while low contrast scenes would self-tune
to slower frames for the same amount of contrast information. Other more sophisticated options
could use a post processing event based system for performing a given recognition and provide
the reset pulse once a recognition has been achieved, or reset after a preset time if no recognition
was possible. In what follows we count a fixed number of events M. Figure 4.25 illustrates the
effect of changing M when observing the paper clip of Figure 4.24. Note that setting M to low values also removes background noise.
The TFS output mode is also insensitive to illumination (in first order), since it operates
directly on Icont within the integrate-and-fire circuit (see Figure 4.9(a-b)). To show this, several
snapshots of the paper clip of Figure 4.24 were taken under different illumination conditions. As
shown in Figure 4.26, Tframe is the sum of Tfirst (the time the retina needs to generate the first
spike after the reset) and TM (the time between the first and M-th spike). ,Figure 4.27 shows the
value of Tframe for different values of M and illumination levels. Tframe is almost independent on
illumination and is approximately constant for a given M. Figure 4.27 also shows the value of
Tfirst versus illumination. In principle, Tfirst should not depend on ambient light because this reset
is performed within the integrate-and-fire circuit (see Figure 4.9(a)) and not the photo sensing cir-
132
Chapter 4: The AER Spatial Contrast Vision Sensor
−2
Tframe&Tfirst (sec)
10
−3
10
−4
10
M=600
M=500
M=400
M=250
T
first
1
10
2
10
3
10
Illumination (lux)
4
10
5
10
Fig 4.27: Effect of Illumination on Tframe and Tfirst
cuit (Figure 4.8). However, Figure 4.27 reveals a slow-down process when decreasing ambient
light (between 5k-lux and 200 lux, approximately). This is probably due to switching crosstalk
between the integrate-and-fire and photo sensing circuits, which introduces a switching transient
in the latter that cannot be prevented when the photo currents are too small. Such problem can be
attenuated in future designs by improving decoupling between the two stages, for example,
through cascoding techniques.
4.5.9 Power Consumption
Chip power consumption has been characterized. Supply voltage is 3.3V. In principle, it
would depend on both static bias conditions and output event rate. However, in practice, it is dominated by the latter, because of the high consumption of digital pads communicating output
events. Static power dissipation is negligible, since pixel current biases are set to relatively low
values. Typical bias settings are I u = 150pA , I low = 50pA and I high = 50pA . This results
in a pixel static current consumption of 15nA. At very low output event rate (1keps) we measured
a chip current consumption of 40µA ( 130µW ). Figure 4.28 shows the measured current consumption of the chip as a function of output event rate. As can be seen, for normal operation
regimes (between 100keps and 1Meps) current consumption varies between 200µA and 2mA
( 660µW - 6,6mW ).
133
Chapter 4: The AER Spatial Contrast Vision Sensor
−2
Current Consumption (A)
10
−3
10
−4
10
−5
10
3
10
4
10
5
6
10
10
Output Event Rate (e.p.s.)
7
10
Fig 4.28: Chip total current consumption as function of total output event rate
Pixel output frequency (or TFS timing) range is directly controlled by bias current Iu (see
Figure 4.8 and Figure 4.9(a)). Therefore, Iu controls also the overall power consumption and the
speed-power trade-off.
4.6 Discussion
In this chapter, a new AER signed spatial contrast retina has been presented. It uses an
improved and calibrated version of Boahen’s contrast circuit. The design avoids the problem of
AER communication bandwidth consumption present in prior designs. Furthermore, it also
includes a thresholding mechanism, so that only pixels sensing spatial contrast above a given
threshold generate events. A calibration scheme is included to partially compensate for pixel mismatch. An optional TFS coding scheme is also available. Extensive experimental results from a
test prototype of 32 x 32 pixels, fabricated in a 0.35µm CMOS technology, are provided. Table
4.3 summarizes the imager main specifications.
An interesting advantage of this contrast retina is its fast response time as well as low
communication throughput, compared to commercial video cameras rendering full frames every
30-40ms. Information throughput is reduced because only relevant contrast information is
provided. Regarding speed response, for example when operating in rate coded mode, since active
pixels fire at frequencies in the range of 1-5KHz, they would all update its state within fractions of
134
Chapter 4: The AER Spatial Contrast Vision Sensor
one mili second, independent of ambient light. In TFS mode, the first front of relevant events
( M = 250 in Figure 4.27) is available in less than 1ms. If the stimulus changes, the retina latency
depends on lighting conditions, ranging from about 100ms at sun light (50k-lux) to 10ms at moon
light (2 lux), with 1ms for indoor ambient light (1 k-lux).
Consequently, the complexity of developing spike based AER spatial contrast retinae, as
opposed to conventional frame-scanned video cameras, could justified by its higher speed
response for a very wide range of illumination conditions, while maintaining the information
throughput low and ambient light independent. Although information throughput is low, relevant
(contrast) information is preserved, which results in significant processing performance
improvement for subsequent stages.
Table 4.3: Sensor Specifications.
Function
ality
Spatial
Contrast to
Number
of
Events
Light to
time
restriction
Latency
NO
0.1-10ms
Dynamic
Range
FPN
>100dB
0.6%
135
Fill
Factor
2%
Pixel
Size
µm
2
80x80
Fabrication
Process
Power
0.35 µm
4M 2P
0.6666mW
CHAPTER 5
The AER Temporal Contrast Vision Sensor
5.1 Introduction
Motion detection is one of the most important tasks that the brain can perform. By detecting
movement, plenty of relevant information about our environment is received. Temporal contrast
detection is useful for the perception of relative depths, 3D structures, and for grouping objects
that we can see. The motion perception pathway is independent of other visual pathways [58].
AER imagers are inherently faster than frame-based sensors. The main reason is that they
can detect illumination changes continuously. There are not dead times between frames where
changes can not be detected.
Several AER temporal contrast vision sensors have been reported previously. The first
attempt to design a temporal contrast sensor was done by Mallik et al. [7] in 2005 and was based
on the prior work of V. Gruev et al. [48] describing a pipeline temporal difference imager. Mallik’s retina modified the traditional active pixel sensor (APS) to detect a quantized absolute
change in illumination. It was a frame-based device with some advantages of AER systems. Several motion detection sensors have been reported [9]-[24] recently. They are not frame-based
devices and have several advantages over conventional vision systems based on frames. Their
most important features are high dynamic range, very high speed tracking and very low power
consumption.The first specific temporal contrast sensor was proposed by Jörg Kramer [59]-[60].
It was a combination of Tobi Delbück’s adaptive photoreceptor [61] with a rectifying and thresholding differentiating element [62]. After this, Lichsteiner et al. proposed an improved temporal
contrast sensor [10]-[9] based on the Kramer’s optical transient sensor. It has high dynamic range
(120dB), low latency (15 µs ), good fill factor (8.1%), and low noise (2.1% contrast). Its bandwidth was limited by the photoreceptor bandwidth.
In this chapter, we present a new AER temporal vision sensor. It is based on the prior work
of Lichtsteiner et al. [9]. Its pixels respond asynchronously to relative changes in intensity. The
sensor directly encodes scene reflectance changes reducing data redundancy and bandwidth consumption. The new imager has improved features and pixels have smaller size. As we will discuss
later, it is specially useful for very high speed tracking and surveillance applications.
In order to increase the sensitivity and the speed response of the retina, two amplifiers were
added after the photoreptor stage. Both of them operate in strong inversion. By this way, the total
137
Chapter 5: The AER Temporal Contrast Vision Sensor
Photoreceptor
Vbp1
Mp
Mbp1
G1
A1
Mcas Vbn1
G2
Vbn1
A2
Vcap
Communication
interface
Vrfr
Vref
Vg
Vcas
Differencing circuit
Amplifiers
Ibn
5C
Vdiff
Rqst
Communication
Interface
C
Vth+
Mm
Vldc1
Ack
Vth-
Vldc2
Reset
Fig 5.1: Overall block diagram of the contrast pixel
gain of the pixel can be increased over the prior design proposed by Lichsteiner and his colleagues. In that design, the total pixel gain was equal to the ratio between the two capacitances at
the differencing stage. To achieve a total gain of 20 with good matching, large capacitors were
required. If we use amplifiers, we can decrease the gain of the differencing circuit. This means
that we can reduce the size of the capacitors and achieve a higher fill factor and higher gain at the
same time. More gain also implies more bandwidth and more sensitivity to temporal contrast
stimuli. For high speed applications, high gain and sensitivity are desirable because changes in
local illuminance occur very fast. As we will discuss later, high gain also implies low latency
response. The only drawback of using amplifiers is more power consumption. However, as we
will discuss later, the total chip consumption has reasonable values in comparison to prior
designs.
The new imager was fabricated and tested. A 128 x 128 pixels prototype was designed in a
0.35µm CMOS technology. Extensive experimental results are provided showing the feasibility
of the design. Examples, of high speed tracking applications are also displayed in the chapter.
5.2 The Pixel Circuit
Figure 5.1 shows the overall diagram of the temporal contrast pixel. The pixel contains four
main parts: 1) The photoreceptor, 2) Two amplifiers to increase sensitivity and speed response, 3)
A differencing circuit to detect temporal luminance variations, 4) Communication interface with
the periphery.
138
Chapter 5: The AER Temporal Contrast Vision Sensor
Vin
M1
M2
Vout
V1
M3
Ibn1
Vldc
Fig 5.2: Schematic of the amplifiers.
The two amplifiers are connected to a external circuit (AGC control) that sets the gain of
amplifiers to a value that does not depend on illumination.
Let us describe all these blocks that compound the pixel circuit and the external AGC circuit.
A.
Photoreceptor
The first stage is the photoreptor circuit. That circuit responds logaritmically to changes in
local illumination. There is an offset at the output of this stage, Vout, that will be amplified later.
This is not a problem because it will be removed with the differencing circuit than only responds
to variations in the local illumination. By adjusting Vg, Vbp1 and Vcas, we can vary the speed of the
photoreptor. The voltage at the output is given by the following expression,
I ph
I ph
V out = κ p V g + nU T ln  ------ = K + nU T ln  ------
I sp
I sp
B.
(5.1)
Pre-amplifiers
To increase the sensitivity and the speed response of the retina, two amplifiers were added
after the photoreptor stage. Both of them operate in saturation region. If we compare the new
design to the prior design of Lichsteiner [9], the total gain has been increased and the pixel area
has been reduced to the half. As we will discuss in section 5.4.1, FPN is similar because the ratio
between gain and interpixel dispersion is approximately the same.
Figure 5.2 displays the schematics of one amplifier. Transistors M2 and M3 operate in strong
inversion. The value of Vldc is set by an external AGC circuit described in section 5.3. By this
way, DC levels at the output do not depend on illumination. In Figure 5.1, we can see how amplifiers are connected to the AGC circuit. Ibn1 is a current source that generates a bias current exter-
139
Chapter 5: The AER Temporal Contrast Vision Sensor
nally. M1 is a source follower that introduces a voltage shift in Vin that can adjusted varying Ibn1,
so V 1 = V in – ∆ V . If M2 and M3 are above threshold saturation region,
W
W
------2- β n2 ( V 1 – V ldc – V Th ) 2 =  ------3- β p3 ( V dd – V out – V Tp ) 2
 L2 
 L3 
(5.2)
Solving for Vout, and knowing that V 1 = V in – ∆ V ,
( W 2 ⁄ L 2 )β 2
V out = V dd – V Tp – --------------------------- ( V – ∆V – V Th – V ldc )
( W 3 ⁄ L 3 )β 3 in
(5.3)
Where ∆V can be set by adjusting Ibn1. For a good operation, ∆V ≈ 0.7V . There are two amplifiers (A1 and A2) in each pixel, so the total gain at the output of the amplification stage is
GT = G1 G2 =
( W 2 1 ⁄ L 2 )β 21 ( W 22 ⁄ L 2 )β 22
1
2
-------------------------------- -------------------------------( W 3 1 ⁄ L 3 )β 31 ( W 32 ⁄ L 3 )β 32
1
(5.4)
2
In our particular case, both amplifiers were designed to have the same gain G 1 ≈ G 2 ≈ 4 and
∆V cap = G 1 G 2 ∆V out . The DC component at the output will be removed with the differencing
circuit. Gain mismatch is low and mismatch because amplifiers transistors operate in strong inversion and there is a good matching between capacitors. Mismatch at the DC component does not
affect the pixels output. Transistor sizes were chosen to operate in strong inversion with reduced
currents (100nA-400nA). Thus, power consumption was reduced. Amplifiers are connected to the
AGC block. That block is placed in the periphery and senses the average photo current. This
block sets the values of Ildc1 and Ildc2, according the global illumination level. This way, DC levels at the amplifiers output are controlled and kept within the amplifiers dynamic range.
 W2 
0,6µm  W 3 
2,4µm  W 2 
0,6µm
The transistor aspect ratios were  ---------1 = ---------------- ,  ---------1 = ---------------- ,  ---------2 = ---------------16µm  L 31 
2,4µm  W 22
8µm
 L 21 
 W3 
1,2µm
and  ---------2 = ---------------- .
1,2µm
L
 32 
C.
Differencing Circuit
This stage responds to relative changes at the voltage input with a gain equal to
G 3 = 5C ⁄ C . There is a good match between the two capacitors. The differencing circuit also
removes the DC component at its input. We can express the variations of voltage at the output as
∆V diff = – G 3 ∆V cap = – G 1 G 2 G 3 ∆V out
140
(5.5)
Chapter 5: The AER Temporal Contrast Vision Sensor
And from eq. (5.1),
I ph ( ∆t + t )
∆V out = nU T ln  -------------------------
 I ph ( t ) 
(5.6)
The total pixel from the input to the output of this stage is G t = G 1 G 2 G 3 .
D.
Communication Block Circuit
Figure 5.3(a) shows the schematics of the communication circuit block. Pixels can detect the
sign of the temporal contrast. There are two comparators with different adjustable thresholds Vhigh
and Vlow that generate positive or negative events (signed contrast) when Vdiff is above the positive threshold or below the negative threshold respectively. There is also a refractory circuit (see
Figure 5.3(b)) that allows to control the time between consecutive events that one individual pixel
can generate. This is specially useful when there are pixels with very high activity and we want to
control the maximum activity in the AER bus. Finally there are two identical communication
blocks, see Figure 5.3(c), that interact with the arbiter and the periphery. These are standard AER
pixel communication circuits taken from Boahen’s row parallel event read-out technique [45].
When generating signed events, each pixel needs to provide two column event signals col+ and
col-. This concept was already implemented and tested in the AER spatial contrast retina [23] and
prior designs [27] that required signed events.
5.3 Automatic Gain Control Block
This block is placed at the chip periphery and allows to control the DC levels at the output of
the two amplifiers adapting them to the global chip illuminance level. In Figure 5.4(a), we can see
the schematics of the AGC circuitry. Its function is to maintain constant the DC levels at the output of the amplifiers of Figure 5.1, A1 and A2, by adjusting the voltages Vldc1 and Vldc2. All pixel
amplifiers, A1 and A2, are connected to the external nodes, Vldc1 and Vldc2, respectively. The AGC
control block has two identical amplifiers to the ones of Figure 5.2. The outputs of these amplifiers are set to a constant voltage, Vodc1 and Vodc2, that we can adjust. To set these constant voltages,
there is one operational amplifier connected to the output of each amplifier. Both of them have
negative feedback to set up the voltages Vldc1 and Vldc2 that will be identical at the amplification
stage and the AGC block.
The input of the circuit of Figure 5.4(a) is the summed photo current sensed by 128 photo
diodes placed around the pixels matrix, by this way voltages Vldc1 and Vldc2 can be set according
to the overall chip illumination level. In Figure 5.4(b) we can see the circuit that senses the average photo current level. It has 128 photo diodes (distributed along the chip periphery) connected
to one photoreceptor identical to the one of Figure 5.1. In this circuit, there is also a femtoampere
current mirror [36] that copies the average sensed photo current. Finally, the output of circuit of
Figure 5.4(b) is connected to the input of the AGC as is indicated in Figure 5.4. There is a capacitor at this input node to make the AGC block insensitive to very quick changes at the luminance
level. We are only interested in slow global changes in the average chip illumination. The value of
141
Chapter 5: The AER Temporal Contrast Vision Sensor
col+
(a)
Vhigh
col-
pulse+
Vrfr
+ event block
Vdiff
Vref
ev_rst+
Ack
Refractory
Circuit
+
ev_rst
ev_rst-
ev_rstpulse- - event block
Vlow
(c)
Vdiff
(b)
Rqst
Crfr
col
Vrfr
Rqst
ev_rst-
pulse
ev_rst+
ev_rst
Ack
Vref
Fig 5.3: (a) Integrate and fire neuron with AER pixel intercommunication modules. (b)
Detail of the refractory circuit. (c) Detail of AER pixel communication blocks.
this capacitor was chosen to have a time constant of milliseconds. This is enough to detect a global change in illumination. For example, when we go out of a house, and the average photo current value, I ph , changes significantly.
142
Chapter 5: The AER Temporal Contrast Vision Sensor
Vldc2
Vldc1
(a)
Vbpn1
A2
A1
Vg
Vbn1
Vcas
Vodc1
Vbn1
Vodc2
I ph
C1
C3
C2
(b)
Vbrf
Ibrf
Vg
Vg
I ph
Vcas
Photoreceptor
Ilcm
I ph1
I ph128
1
Photo diodes
gnd01
Vdd01
Femtoamper
current mirror
Fig 5.4: (a) AGC circuit block. (b) Detail of the circuitry used to sense the average chip
illuminance, I ph , highlighting its main parts.
143
Chapter 5: The AER Temporal Contrast Vision Sensor
Table 5.1: Chip Specifications.
technology
power supply
chip size
array size
pixel size
fill factor
photodiode quantum
efficiency
pixel complexity
current consumption
CMOS 0.35µm 4M 2P
3.3V
5.54 x 5.69 mm2
128 x 128
35.7 x 35.5 µm2
8%
0.34 @ 450nm
58 transistors + 3 caps
44mA @ 10keps
5.4 Experimental Results
A 128 x 128 pixel AER signed temporal retina chip was designed and fabricated in a double
poly 4-metal 0.35 µ m CMOS process with a power supply of V DD = 3.3 V. Table 5.1 summarizes the chip specifications. Figure 5.5 shows a micro photograph of the die, of size
5.689x5.538mm2. The whole chip, except the pad ring, is covered with the top metal layer leaving
openings for the photo diode sensors. Figure 5.5 also shows the layout of a single pixel highlighting its main components. Each pixel layout is a symmetrical reflection of its neighboring pixels.
This way noisy digital lines are shared among neighbors, as well as power supplies, and noise
sensitive bias lines. At the same time, noise sensitive lines are separated from noisy ones. Pixel
2
area is 35.5x35.7 µm , including routing. The pixel was made up of 58 transistors and 3 capacitors (the two of the capacitive divider of the differencing circuit and the capacitance of the refractory circuit).
Figure 5.6 shows the experimental setup to characterize the temporal contrast sensor. The
chip was mounted on the Caviar Board (a specific PCB designed to test generic AER devices).
This board was connected to a USB port. The digital words to program the bias currents of each Ipot [35] were sent through this port. The outputs of the Caviar Board were sent to an AER bus.
This bus was connected to a Mapper. The Mapper was a programmable device that was used to
filter or remove the activity of some pixels for some specific measurements (bandwidth and
latency characterization). The output bus of the Mapper was connected the board USB2AER [46],
which sends the AER data from the bus to one of the USB computer ports. This information was
processed by a jAER [47] that allows to see real time images. The USB2AER board was specially
useful for real-time monitoring. The input bus to this board was replicated and sent to a Datalogger. This board has an internal memory that can store the data generated by 524,000 events and
was used to save efficiently on the PC the AER information transmitted from the AER bus. To
144
Chapter 5: The AER Temporal Contrast Vision Sensor
5.538mm
bias
generators
5.689mm
128x128 pixel array
AER out
Global
AGC
AER communication block
refractory
capacitor
capacitive divider
35.5 µ m
photo diode
amplifiers
comparator
35.7 µ m
Fig 5.5: Microphotograph of 5.538mm x 5.689mm die, and zoom out of 35.7 µ m x 35.5 µ m
pixel (layout) indicating the location of its main components.
145
Chapter 5: The AER Temporal Contrast Vision Sensor
usb
Mapper
USB2AER
AER BUS
usb
AER BUS
Datalogger
STIMULUS
Ipots
Caviar Board
usb
Datalogger
Lens Mount
Mapper
Caviar Board
USB2AER
Fig 5.6: Experimental setup and photo of the retina and main boards used to test the
imager.
146
Chapter 5: The AER Temporal Contrast Vision Sensor
generate programable stimuli, a monitor can be placed over the retina visual field. The stimuli can
be generated with a Matlab routine.
The nominal bias settings to test the retina were V ref = 1.5V , V g = 600mV ,
V cas = 1.8V , V bpn1 = 2V , I bn = 100nA , I bldc = 100nA , I brf = 10nA and I bn 1 = 200nA .
As we will discuss in section 5.4.4, the bias current I bn 1 can be adjusted to control the pixels
bandwidth and the amplifiers power consumption.Voltages were generated with external potenciometers and bias currents were generated by using the I-Pots described in chapter 2.
5.4.1 Uniformity of Response
One important feature of imager sensors [10]-[9] is the FPN that characterizes the uniformity of the response. In our particular case, we were interested in knowing how was the pixel contrast threshold variation, θ ev , to generate one event. Ideally, all the pixels should generate the
same number of events when they are stimulated with the same stimulus. For one individual pixel,
we can define the temporal contrast as
I ( t + ∆t ) – I ( t )
θ = 2 ⋅ -----------------------------------I ( t ) + I ( t + ∆t )
(5.7)
Where I ( t ) is pixel photo current value at one instant, t. If both photo currents are very similar, I ( t ) ≈ I ( t + ∆t ) , then we can approximate the temporal contrast by
I ( t + ∆t ) – I ( t )
∆t
θ ≈ ----------------------------------- = -------I(t)
I(t)
(5.8)
where ∆t = I ( t + ∆t ) – I ( t ) .
From eq. (5.1), we know that a given change in the pixel photo current will produce the following change in Vout,
I ( t + ∆t )
∆V out = nU t ln  --------------------
I(t)
(5.9)
Combining eq. (5.5) and eq. (5.9), and knowing that the total gain at the output of the differenciator is G t = G 1 G 2 G 3 (where G 1 , G 2 and G 3 are the gains of the amplifiers and differentiator, respectively), we can express the change at the differencing circuit output ∆V diff as
I ( ∆t + t )
∆V diff = – G t U t n ln  -------------------- = – G t U t n ln ( θ + 1 )
 I(t) 
(5.10)
And usually θ « 1 , so ∆V diff is approximately given by the following expression
∆V diff = – G t U t nθ
147
(5.11)
Chapter 5: The AER Temporal Contrast Vision Sensor
After the differencing circuit, there are two comparators that generate positive or negative
events (see Figure 5.3(a)) depending on the value of Vdiff and the positive and negative thresholds,
V low and V high . If we set symmetric thresholds, one positive or negative event will be generated
when there is a change in Vdiff, ∆V diff_ev , and
– ∆V diff_ev = V high – V ref for a positive event
(5.12)
– ∆V diff_ev = V ref – V low for a negative event
(5.13)
Combining eq. (5.12), (5.13) and (5.11), we can obtain an expression for the minimum temporal contrast, θ ev , necessary to produce one single event (either positive or negative)
∆V diff_ev
θ ev = ± -----------------------nG t U t
(5.14)
And when there is a slow temporal change in the photo current, if the output voltage variation is
∆V diff
higher than ∆V diff_ev , the pixel will generate N = --------------------- events. Hence,
∆V diff_ev
∆V diff_ev
ln ( ( I ( ∆t + t ) ) ⁄ I ( t ) )
θ ev = ± ------------------------ = ------------------------------------------------nG t U t
N
(5.15)
In order to characterize the FPN, we stimulated the retina pixels with a black bar with linear
gradient edges that was moving at constant speed through the visual field of the sensor [9]. The
bar was passed through the visual field 30 times. After this, the average number of events generated by each pixel, N , was recorded. Ideally, if all the pixels are stimulated with the same stimulus, they should spike the same number of times. The goal of the experiment was to quantify the
pixel deviation. The effect of the refractory period was reduced to the minimum. The experiment
was also repeated for different positive and negative symmetric thresholds,
∆V diff_ev = [ 0.2, 0.24, 0.33, 0.33, 0.41, 0.46, 0.51 ]V . The stimulus was generated with a LCD
monitor. Its contrast and speed was controlled with a computer program. The number of events
does not depend on the speed of the stimulus. It only depends on its contrast. In our particular
case, the ratio between the photo currents measured at the brightest and the darkest grey levels of
the stimulus was I bright ⁄ I dark ≈ 4 . We determined this ratio by measuring the reflectance of the
brightest and the darkest grey levels. Let us define the log contrast as
I bright
C = ln  --------------
 I dark 
Therefore, an approximation for the contrast event threshold of eq. (5.15) is
148
(5.16)
Chapter 5: The AER Temporal Contrast Vision Sensor
C
θ ev = ---N
(5.17)
Figure 5.7 shows the histograms with the average number of positive and negative events
generated for each pixel of the retina. These values of the contrast event thresholds were calculated using eq. (5.17) and are indicated on each graph. For the highest threshold, there is an average of 3.6 positive and 3.6 negative events per positive and negative edge. For the lowest
threshold, there is an average of 16.1 events per pixel (either positive or negative events). For
lower thresholds, the average number of events and the deviation is higher.
5.4.2 Pixel Gain Measurements
From the experimental results of Figure 5.7, we can obtain the average gain of each pixel.
Combining eq. (5.15) and eq. (5.11), we can express the total gain (number of events per log contrast of the stimulus) as
N ⋅ ∆V diff_ev
G t = -----------------------------n ⋅ Ut ⋅ C
(5.18)
Figure 5.8 displays the average measured individual pixel gain values and the number of
pixels which have each gain value. The estimated gain values are plotted for different values of
the voltage thresholds. For all the pixels, the average gain value is G t ≈ 65 . Ideally speaking,
gain should not depend on the contrast threshold. We can see that for the highest threshold
( ∆V diff_ev = 0.51V , the total gain is approximately 55. The reason is that there are pixels that
do not spike sometimes when the stimulus pass through the visual field and the threshold is too
high. When the threshold is too low, ∆V diff_ev = 0.2V , outliers that spike with very high frequency appear. For this reason, the total gain is slightly higher ( G t ≈ 70 ). For the rest of the
threshold values, the average gain is quite similar, G t ≈ 65 .
5.4.3 Contrast Threshold Deviation
Let us know study the deviation in the contrast threshold from the measurements shown in
Figure 5.7. If there is a variation of ∆V diff higher than the contrast threshold, from equations
(5.10), (5.12) and (5.13), we can obtain the number of spikes that the pixel will generate
∆V diff
ln ( θ + 1 )
N = --------------------- = ----------------------∆V diff_ev
θ ev
(5.19)
Because of pixel mismatch not all generate the same number of events, but there will be a statistical distribution where each pixel will generate a number of events given by
∆θ ev
∆θ ev
ln ( θ + 1 ) ln ( θ + 1 )
N + ∆N = ------------------------- ≈ -----------------------  1 – ----------- = N  1 – -----------

θ ev 
θ ev 
θ ev 
θ ev + ∆θ ev
149
(5.20)
Chapter 5: The AER Temporal Contrast Vision Sensor
450
ON events
OFF events
400
θ ev = 0.49 N = 3.56
350
# pixels
300
θ ev = 0.33 N = 4.33
250
θ ev = 0.26 N = 5.39
200
θ ev = 0.22 N = 6.41
150
θ ev = 0.158 N = 8.97
100
θ ev = 0.125 N = 12.86
50
0
0
5
10
15
# events/ pixel/ edge
20
25
Fig 5.7: Pixel histograms with the average number of events recorded per pass of the stimulus bar for 30 repetitions. The experiment was also repeated for different values of the
event threshold.
150
Chapter 5: The AER Temporal Contrast Vision Sensor
(a)
Average pixel gain
1200
Thresh=0.51V
Thresh=0.46V
Thresh=0.41V
Thresh=0.33V
Thresh=0.24V
Thresh=0.2V
1000
# pixels
800
600
400
200
0
0
20
40
60
Gain
80
100
120
75
Average Gain
(b)
70
65
60
55
0.2
0.25
0.3
0.35
0.4
Voltage Threshold (V)
0.45
0.5
0.55
Fig 5.8: (a) Distributions in the pixels gains obtained from the measurements of Figure 5.7
and eq. (5.18). (b) Average gain values for each voltage threshold.
151
Chapter 5: The AER Temporal Contrast Vision Sensor
10
8
σ θev(%)
6
4
2
0.1
ON
OFF
0.15
0.2
0.25
0.3
0.35
0.4
0.45
θ ev (ln(I) units)
80
ON
OFF
Interpolation
60
σ θ ev
--------- ( % ) 40
N
20
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
θ ev (ln(I) units)
Fig 5.9: Top panel shows the contrast deviation expressed in % as a function of the contrast deviation. Bottom panel shows the ration between the contrast deviation and the
average number of events (also in %).
Consequently,
∆θ ev
∆N
-------- = – ---------N
θ ev
(5.21)
θ ev
σ θev = ------- σ ( ∆N )
N
(5.22)
and
Top panel of Figure 5.9 shows the standard deviation of measured contrast threshold in % (computed from eq. (5.22) and the measurements of Figure 5.7) plotted as a function of contrast thresh-
152
Chapter 5: The AER Temporal Contrast Vision Sensor
σθ
old. At the bottom panel, we have plotted --------ev- as a function of contrast threshold. The minimum
N
experimental mismatch deviation (FPN) was about 2.8%.
The main sources of mismatch are the transistor threshold voltage, ∆V diff , of comparators
and the interpixel mismatch of the gain, G t . If we assume independent statistical distributions, the
contrast threshold can be expressed as
2
2
σ θev = σ θev
2
Gt
+ σ θev
∆V diff
(5.23)
And using the chain rule, we can obtain equations for the two sources of mismatch from eq. (5.15)
σ θev
Gt
σG
∆V diff
= – --------------- = --------t θ ev
nU t G t
Gt
σ θ ev
∆V diff
σ ∆V diff
= -------------nU t G t
(5.24)
(5.25)
And the total threshold deviation can be expressed as follows
2
σ θev
=
2 2
θev σ A
+
2
σo
σ θ ev
σ ∆Vdiff
G
, with σ A = ---------------t and σ o = -------------Gt
nU t G t
(5.26)
In Figure 5.10, we have plotted the probability density function of the contrast threshold to
produce one output event, θ ev , assuming that has a Gaussian distribution. Let us denote as θ evn
the mean value of θ ev for all the retina pixels. Looking at Figure 5.10, if we want the 99% of the
pixels to spike when there is a contrast stimulus, the minimum detectable contrast will be
θ min = θ evn + 3σ θ ev = 6σ θev
(5.27)
We have to highlight that if there were pixels with a value of θ ev = 0 , they would be
always spiking, even if there is no contrast. Therefore, the minimum detectable threshold will be
θ ev = 3σ θev
(5.28)
From equations (5.27) and (5.28), we can state that,
θ min = 2θ ev
(5.29)
And substituting equations (5.27) and (5.29) in (5.26), we can deduce the value of the minimum detectable contrast is
153
Probability
Chapter 5: The AER Temporal Contrast Vision Sensor
– 3 σ θev
θ evn
3σ θev
θ ev
Fig 5.10: Probability density function of the contrast threshold, θ ev .
2σ o
θ min = 2θ ev = -------------------1--2
– σA
9
(5.30)
In our particular case, we can compute the values of σ o and σ Gt from the experimental
results of Figure 5.9 and substitute them in eq (5.26) to obtain the minimum detectable contrast
θ min . Note that in (5.26), for low values of θ ev , σ θ ev ≈ σ o , because mismatch at the offset of
comparators is the main source of mismatch. Therefore, σ θ ev would be approximately equals to
the minimum value of the curve at top panel of Figure 5.8. For higher values of θ ev ,
σ θev ≈ θ ev σ A . Thus, σ θev is approximately the slope of the curve at top panel of Figure 5.8. In
this case, amplifiers gain is the main source of mismatch. Looking at Figure 5.8, σ A ≈ 0.16 and
σ o ≈ 0.025 . Therefore, the minimum contrast that the 99% of the pixels can detect is
2σ o
θ min = --------------------- ≈ 16.9%
1--2
– σA
9
154
(5.31)
Chapter 5: The AER Temporal Contrast Vision Sensor
And the comparators offset is given by
σ Voffset = σ ( ∆V diff ) = G t U t nσ θev
σG = 0
= G t U t nσ o = 30mV
(5.32)
t
According to eq. (5.27), ∆V diff has to satisfy ∆V diff > 6 ⋅ 30 = 180mV . On the contrary,
some pixels will be always spiking. Experimentally, we could only set values of ∆V diff above
0.2V. For lower values of voltage thresholds, the event rate was too high (there were pixels spiking all the time with θ ev = 0 ) and the arbiter was not able to send out all the events properly.
5.4.4 Bandwidth Measurements
One of the more impressive features of the sensor is the pixel bandwidth. If we calculate the
transfer functions of the main subsystems that compound the pixel (photoreceptor, amplifiers and
differencing circuit), we can understand the dynamic response of retina pixels. Let us study all
these blocks separately to analyze their influence on the dynamic response and obtain their transfer functions. The overall transfer function will be equal to the product of these individual transfer
functions.
As we will show, the photoreceptor limits the pixel bandwidth. In Figure 5.11(a), the photoreceptor and its main parasitic capacitances are shown. If we calculate the transfer function
between the input and output nodes (V in and V out ), taking into account the main parasitic capacitances, we obtain a second order low-pass filter with a transfer function given by the following
expression:
g mm g mcas
g mcas g mm
1
- = – ------------------- ⋅ ------------------------------------------------------,
H 1 ( s ) = – ------------------------------------------------------------------2
Cd C1
g mp g mcas
2 g mp C 2
C d C 1 s + g mp C 2 s + g mcas g mp
s + -------------- s + ------------------C1 Cd
C1 Cd
(5.33)
with C 1 = C gsp + C gdbp1 + C gd cas
(5.34)
and C 2 = C gscas + C ds m + C d
(5.35)
If we calculate the poles of eq. (5.33),
g mp C 2 1
- ± --- ⋅ g 2mp C 22 – 4g m g m C 1 C d
s 12 = – --------------p
cas
2C 1 C d 2
(5.36)
Under typical operation conditions, g mp < g mcas because transistor Mg operates in weak inversion
2
2
and transistor Mcas operates in strong inversion. Hence, g mp C 2 < 4g mp g mcas C 1 C d and poles are
complex conjugates. However, there might be situations were poles were real number, depending
of the chip illumination and bias settings. A generic second-order low-pass filter with complex
conjugates poles can be expressed as,
155
Chapter 5: The AER Temporal Contrast Vision Sensor
(a)
Vbp1
C gs
p
C
Mp
C gdbp1
C
(b)
cas
M3
M2
C ds1 C ds2 C gs4
M4
(c)
I bn
M1
V out
5C
C ds3 C ds
4
C gs2
C gs1
V in
Vout
C gd 4
C gs1
V ref
ds m
C gs3
M1
Vbn1
C
Mm
m
C gd1
Vin
Mcas
Vcas
gd p
C gs
C gs
Vout
gd cas
Vg
Vin
Cd
Mbp1
C gd3
C
V ref
M2
C ds 3
C ds4
V out
C = C2
M4
M3
V reset
C gd3
C ds2
Ib
C gs4
5C = C 1
V in
Fig 5.11: Schematics of the main pixel blocks used to calculate the pixel transfer function. The main parasitic capacitances have been highlighted in each schematic. (a) Photoreceptor, (b) Amplifiers, and (c) Differencing circuit with circuitry detail including the
amplifier and its main parasitic capacitances.
156
Chapter 5: The AER Temporal Contrast Vision Sensor
k
H ( s ) = --------------------------------2 wo
2
s + ------ s + w o
Q
(5.37)
We can obtain the filter cut-off frequency, assuming complex poles,
wo1 =
g mcas g m
-------------------p
Cd C1
(5.38)
Studying eq. (5.38), we can state that pixel bandwidth depends of transistor M m conductance. Transistor M m works in weak inversion. Conductance changes with illumination because,
I ph
g mp ≈ ------- . For this reason, pixel bandwidth is proportional to
UT
I ph . For low luminance levels,
g mp has low values and pixel bandwidth is lower. We also have to highlight that C d , the photo
diode parasitic capacitance has a higher value than other parasitic capacitances. For this two reasons, this stage limits pixel bandwidth. Figure 5.12 displays the bode diagram of the photoreceptor transfer function. We can also see how bandwidth changes when we vary illumination
conditions. The phase is approximately flat at the frequencies of interest and equal to -90º because
the sign of the output voltage is inverted as is depicted in eq. (5.33).
In the amplification stage, we find two identical amplifiers. In Figure 5.11(b), there we can
see the schematics of one of them. The main parasitic capacitances are indicated in the figure. The
current source of Figure 5.2 has been replaced by transistor M2 in Figure 5.11(b). This transistor is
part of a current mirror that feeds the amplifier. The transfer function of each amplifier is approximately given by,
g mp g mp
2 C gs 1 g m 4 – C gd 4 g m 1
s + -------------------– s + ------------------------------------------C gs1 C gd4
C gs1 C gd4
H ( s ) = – ---------------------------------------------------------------------------------------------------g m1 g m3
2 g m 3 C 3 + g m 4 C gd m + g m 1 C 4
- s + --------------s + ------------------------------------------------------------2
2
C5
C5
2
(5.39)
With C 5 = C gd 4 ( C 1 + C gs1 ) + C 2 ( C gs1 + C gd 4 + C 1 ) ,
(5.40)
C 4 = C gd4 + C 2
(5.41)
C 3 = C gs1 + C gd4 + C 1
(5.42)
C 2 = C ds3 + C ds4
(5.43)
157
Chapter 5: The AER Temporal Contrast Vision Sensor
Photoreceptor Bode Diagram
50
Iph
Magnitude (dB)
0
−50
-40dB/dec
−100
−150
−200
0
Iph
Phase (deg)
−45
−90
−135
−180
0
10
2
4
10
10
6
10
8
10
Frequency (Hz)
Fig 5.12: Bode diagram of the photoreceptor transfer function. It is shown how illumination affects pixel bandwidth.
C 1 = C ds1 + C ds2 + C gs4
(5.44)
Let us denote τ = g m3 C 3 + g m4 C gd m + g m1 C 4 , the poles of eq. (5.39) are equal to the following expression
τ
1
2
2
s 12 = – ------2 ± ---------2- ⋅ τ – 4g m3 g m1 C 5
C 5 2C 5
(5.45)
2
2
Assuming that g m1 ≈ g m3 ≈ g m4 , poles are complex conjugates because τ < 4g m3 g m1 C 5
( C 5 is the dominant capacitance). Figure 5.13 shows the bode diagram of the transfer function of
each amplifier. It is approximately a low-pass filter. Amplifiers have a constant gain within the
band pass. For higher frequencies, input signals are attenuated. In this case, the cut-off frequency
is
158
Chapter 5: The AER Temporal Contrast Vision Sensor
Amplifiers Bode Diagram
20
10
Magnitude (dB)
0
−10
−20
−30
fo2
−40
−50
−60
0
Phase (deg)
−45
−90
f o2
−135
−180
Frequency (Hz)
Fig 5.13: Bode diagram of amplifiers transfer function.
wo2 =
g m3 g m1
------------------------------------------------------------------------------------------------ =
C gd 4 ( C 1 + C gs1 ) + C 2 ( C gs1 + C gd 4 + C 1 )
g m3 gm1
---------------2
C5
(5.46)
Amplifiers have been designed to have a bandwidth higher than the photoreceptor stage,
2
w o2 > w o1 . C 5 in eq. (5.46) is lower than the product of capacitances C d C 1 in eq. (5.38). Numerator of eq. (5.46) is also lower than numerator of eq. (5.38) because amplifiers transistors M1 y M3
work in strong inversion. Therefore, we can assume that this stage only amplifies the input signal
coming from the photoreceptor.
Figure 5.11(c) displays the differencing circuit and its main parasitic capacitances. The
amplifier symbol of Figure 5.1 has been replaced by its schematics. In this case, the transfer function is given by the following expression
gm1 g m
gm
2 gm
– s – --------1 s + ----------------4
s + --------1
C3
C3
C2 C3 C1
C3
H 3 ( s ) = – ------ ⋅ ------------------------------------------------- ≈ ------ ⋅ ----------------gm
C4
g m1 gm4 C 2
2 gm
s + --------1 s + --------------s – --------1
C4
C1 C4
C4
159
(5.47)
Chapter 5: The AER Temporal Contrast Vision Sensor
Differentiator Bode Diagram
Magnitude (dB)
150
100
20dB/dec
50
0
180
Phase (deg)
135
90
45
0
−2
10
0
10
2
10
Frequency (Hz)
4
10
6
10
8
10
Fig 5.14: Bode diagram of the differencing circuit transfer function.
C1 and C2 are the external capacitors of the differencing circuit. The ratio between them is
gm gm
equal to the gain of this stage. In our case, G 3 = C 1 ⁄ C 2 ≈ 5 and C 3 > C 4 . Hence, --------1 < --------1 In
C3 C4
Figure 5.14, we can see the bode diagram of its transfer function. The system work as a differentiator from DC to frequencies higher than w o1 . This stage does not limit the pixel bandwidth and
also introduces and additional gain.
In conclusion, the logarithmic photoreceptor limits pixel bandwidth. This depends on illumination because the conductance of transistor Mp in Figure 5.1 is proportional to illumination,
I ph
g mp = ------- . We can consider pixels as a second-order systems. The total transfer function is given
UT
by the product of the three transfer functions, H ( s ) = H 1 ( s )H 2 ( s )H 3 ( s ) . It can be approximated
as H ( s ) ≈ G 1 G 2 G 3 H 1 ( s ) = G t H 1 ( s ) , where G1 and G2 are the amplifiers gain and G3 is the differencing circuit gain.
The photoreceptor has been chosen to improve the pixel bandwidth. In the prior design of
Lichsteiner [9], there was a Miller capacitance between the input and the output node. With the
new configuration, this limitation has been solved.
160
Chapter 5: The AER Temporal Contrast Vision Sensor
R = 1kΩ
V in ( t ) = V offset + A ⋅ sin ( 2πft )
Fig 5.15: LED modulated with a sinosoidal wave used to take the bandwidth measurements.
To take the bandwidth measurements, a group of pixels of the center of the matrix were
stimulated with a LED. The diode was modulated with a sinosoidal signal (see Figure 5.15). The
signal has an offset level to keep the LED always ON, so V offset – A > V γ . We were varying the
modulation frequency and counting the number of positive and negative events generate per stimulus cycle during a scan period of 20s. In the band pass, this ratio remains constant. For frequencies higher than the cut-off frequency, this ratio becomes smaller. In Figure 5.16, there are some
transfer functions of several pixels. The x-axis represents the stimulus frequency and the y-axis
the number of events (positive and negative) per cycle of the sinewave. Measurements were taken
at 1Klux with indoor light. Some pixels even exhibit a small resonant peak about 2KHz. There is a
mismatch between the pixels bandwidth. All the pixel transfer functions have a bandwidth higher
than 5KHz. Some pixels display a bandwidth even higher than 10KHz. The average pixel bandwidth was about 8KHz. This gives an idea about how fast can be the temporal changes that the retina can detect. We set symmetric positive and negative thresholds to take the measurements.
Figure 5.17 depicts how the pixel bandwidth depends on illumination. As we explained previously, pixel bandwidth is proportional to photo current when transistor Mp is working in weak
inversion. In Figure 5.17, initially, illumination was 1Klux (for higher values of illumination, the
shape of the pixel transfer functions does not change significantly and pixel bandwidth remains
constant) and pixel bandwidth was about 6KHz. Then, the event transfer function was measured
several times by using neutral density filters to decrease the illumination level. Bandwidth is proportional to illumination. For low illumination levels (10 lux), the pixel still has a bandwidth of
300Hz that is enough to detect very quick changes of illumination in the scene. In Figure 5.17, top
panel shows experimental event transfer functions for different illumination levels. Pixels have
approximately a first-order response. Fitting measured data with first-order transfer functions, we
achieve a good matching for all the illumination values. We have plotted in red the first-order
transfer functions that fit the data of the curves measured with lowest and highest illumination
levels. For the rest of the curves, there is also a good matching with first-order transfer functions.
Bottom panel displays how pixel bandwidth depends on illumination. We have plotted pixel bandwidth as a function of illumination. The linear fitting of the measurements is also displayed. For
illumination levels higher than 1Klux, pixel bandwidth remains constant and does not change significantly. Ideally speaking, amplifiers gain should be the same within the band-pass. However,
there are small variations within the band-pass of the number of events per cycle for each value of
illumination. These variation could be due to a non perfect tune of the external voltage Vldc.
161
Chapter 5: The AER Temporal Contrast Vision Sensor
Several Event Transfer Functions
4
3.5
# Events/Cycle
3
2.5
2
1.5
1
0.5
0
0
10
1
10
2
3
10
10
Stimulus frequency (Hz)
4
10
5
10
Fig 5.16: Several event transfer functions obtained when a group of pixel of the center of
the retina were stimulated with a LED, as is shown in Figure 5.15.
It is possible to exert some control over the pixels bandwidth and the power consumption.
Obviously, the more bandwidth the more consumption. For some applications, it is not necessary
to have a pixel bandwidth higher than 1KHz. Figure 5.18 illustrates how the pixel bandwidth
changes when we vary the bias current I bn 1 that controls the gain of the amplifiers. For bias values higher than 400nA, pixel bandwidth remains constant. For a value of I bn 1 = 80nA , we still
have a pixel bandwidth of 800Hz.
It is possible to control the bandwidth varying other parameters. For example, varying Vg or
V bp1 in the photoreptor. However, it is desirable to keep constant these bias and control the power
consumption and the pixel bandwidth varying the gain of the amplifiers. Its amplifiers work in
saturation region and chip consumption can be reduced significantly by varying amplifiers bias
currents. Bias settings at the photoreceptor has a more restrictive tuning range and bad tuning of
them could affect the pixel response.
In order to characterize the pixel response to very low stimulus frequencies [9], we repeated
the bandwidth results of Figure 5.16 to determine the pixel bandwidth. In this case, we increased
162
Chapter 5: The AER Temporal Contrast Vision Sensor
Dependence of Pixel Bandwidth with Illumination
7
# Events/cycle
1Klux
500lux
100lux
50lux
10lux
5lux
3lux
2lux
1st order fit
2 lux
6
1 Klux
5
4
3
2
1
0
0
10
1
2
10
3
4
10
10
Stimulus Frequency (Hz)
10
5
10
Pixel Bandwidth (Hz)
Measured Bandwidth
Linear Fitting
3
10
1
2
10
10
3
10
Illumination (lux)
Fig 5.17: Top panel shows several transfer functions of one individual pixel when varying
chip illumination. Initially, illumination was about 1Klux. We used neutral density filters to
decrease it. Legend indicates the chip illuminance in lux. Pixels transfer functions are
approximately first-order filters. We have plotted in red the first-order transfer functions
that fit the measured data for the highest and the lowest illumination values. Bottom panel
displays how pixel bandwidth depends on illumination. A linear fitting of the bandwidth
measurements is also displayed.
the scan period to 240s to assure that at least a few events per cycle were generated during the
scan period. The leakage current of the reset transistor in the differencing circuit generates positive events with a frequency of 46mHz. These events have to been take into account when we
study the pixel response to very low frequencies. Figure 5.19 shows the average number of positive and negative events generated by one pixel per stimulus cycle. We have also plotted a third
curve showing how the leakage current of the reset transistor affects the measurements of positive
events. The effective number of positive events per cycle is equal to the difference between the
measured events per cycle and the number of positive events generated due to the leakage current
during the scan period. Looking at Figure 5.19, we can state that the retina can detect changes of
illumination of 10mHz. According to the results of Figure 5.17, the pixel frequency range of oper–2
3
ation is approximately [ 10 , 7 ⋅ 10 ]Hz .
163
Chapter 5: The AER Temporal Contrast Vision Sensor
Bandwidth Control
2.5
I
=400nA
I
=150nA
bn1
bn1
Ibn1=120nA
Ibn1=80nA
2
# Events/Cycle
Ibn1=50nA
1.5
1
0.5
0
0
10
1
10
2
3
10
10
Stimulus Frequency (Hz)
4
10
5
10
Fig 5.18: Bandwidth control varying the bias current I bn1 .
5.4.5 Latency Characterization
As we explained in section 4.5.6, latency is the amount of time that it takes from the occurrence of a illumination change at the photo diode and the output of an address event. This imager
is specially useful for high speed applications an it could be easily used as part of a feedback controller. Latency can be considered as effective measurement of the pixel bandwidth. It also
depends on the illumination level. For low values of illumination, it is reciprocal with illuminance. Figure 5.21 displays the latency measurements of one individual pixel under different illumination conditions. Error bars indicate the deviation (jitter) between different measurements.
Under standard operation conditions, latency is always below 1ms. The lowest experimental value
of the latency was 3.67 µs at 25Klux. Latency measurements were taken under nominal bias settings with I bn 1 = 200nA and ∆V diff = 0.2V . If we set low thresholds, we will reduce the
latency period.
To take latency results, we use a LED (as is shown in Figure 5.15). In this case, the diode
was initially OFF. Then, we illuminated it using a step signal. The latency was the delay between
the step signal and the first RQST of the pixel under test in response to the luminance change. To
164
Chapter 5: The AER Temporal Contrast Vision Sensor
Pixel behaviour at very low stimulus frequencies
3.5
Negative events
Positive events
Effect of leakage
3
# Events/Cycle
2.5
2
1.5
1
0.5
0
−2
10
−1
10
0
10
Stimulus Frequency (Hz)
1
10
Fig 5.19: Pixel response at very low frequencies and influence of leakage current on measured positive events for each frequency.
avoid confusions with the request signals sent out by other pixels, a Mapper device (connected as
is indicated in Figure 5.6) was programed to remove all the Request signals from other pixels. The
delay introduced by the Mapper was known ( ~270ns ). Latency was measured with an oscilloscope working on single mode. Figure 5.21 shows the two recorded signals with the oscilloscope
(step stimulus and Request signal) and the delay (latency) between them. The experiment was
repeated 30 times to determine the pixel deviation (jitter).
5.4.6 Dynamic Range
We define the dynamic range of the sensor as the ratio of the lowest to the highest level of
illumination at which the imager can detect contrast stimuli. The logarithmic sensor compresses
the dynamic range and allows to work under very diverse illumination conditions. We tested the
imager successfully with very illumination ratios under 1lux. It was also able to work under very
bright lighting conditions up to 50Klux. The measured dynamic range was higher than 100dB.
Photo diode dark current limits the lower end of the range. The precision of the photometer used
to measure the level of illumination was 1 lux. To know the exact value of the chip illumination,
we measure overall level of illumination of the room where the sensor was placed (~ 400lux). To
165
Chapter 5: The AER Temporal Contrast Vision Sensor
−1
10
−2
10
−3
Latency (s)
10
−4
10
−5
10
−6
10
0
10
1
10
2
3
10
10
Illuminacion (lux)
4
10
5
10
Fig 5.20: Latency and latency deviation (error bars) versus illumination in response to a
step increase of single pixel illumination. The experiment was repeated 30 times to determine latency deviation. The circuit was fed with nominal bias settings described in section
5.4.
illuminate the chip with lower and controlled levels of illumination, we used neutral density filters.
Figure 5.22 shows two different snapshots. The first one is an image taken under very low
illumination conditions (<1 lux). The second one is a snapshot of an image with two regions with
very different levels of illumination (intra-scene variations). The illumination ratio of the bright
and the dark region was 500. The proposed amplifiers AGC mechanism works well under very
different illumination conditions. We achieved good results in very different natural environments. Its main goal is just to keep the DC gain constant, independently of the global luminance
level.
5.4.7 Example Data and High Speed Applications
Figure 5.23 displays some scenes taken with the temporal contrast sensor. All of them and
the data of Figure 5.24 and Figure 5.26 were captured with indoor light and nominal bias settings
with I bn 1 = 200nA .
166
Chapter 5: The AER Temporal Contrast Vision Sensor
latency
Fig 5.21: Recorded signals used to measure the pixel latency. The delay between the step
stimulus and the first pixel request is the latency. Delay was measured with an oscilloscope.
The sensor is specially suitable for high speed applications. According with the experimental results of Figure 5.16, the imager can detect temporal changes in illumination with a frequency
higher than 10KHz. It is difficult to generate such fast stimuli with mechanical devices. Figure
5.24 shows the events generated during a period of time by a rotating white dot at 400Hz
(24,000rpm). It was painted on the black blade of an electrical fan. The retina easily detected it.
Unfortunately, it was not possible for us to work at faster speed with mechanical devices. In order
to generate faster and controlled stimuli, we used an analog oscilloscope working on XY mode. Its
inputs were two signals with a controlled delay between them. Thus, a bright point rotating at
controlled speed could be generated and seen in the phosphor oscilloscope screen.
Firstly, we stimulated the imager with bright points moving through an elliptical trajectory.
The oscilloscope inputs were two sinosoidal signals with a delay between them. However, we
found out that the phosphor time response could be a limitation to generate high speed stimuli.
Emitted light by the phosphor screen did not go out immediately. There was a time response and it
could occur that one point would be always illuminated if the stimulus was very fast. Using spiral
shape paths, the effect of phosphor persistence was very attenuated because the trajectory of the
stimulus was longer.
167
Chapter 5: The AER Temporal Contrast Vision Sensor
(a)
(b)
32.7ms, 259 events
32.4ms, 304 events
< 1 lux
1 lux
500 lux
Fig 5.22: (a) Captured snapshot under very low illumination conditions (< 1 lux). (b) Captured snapshot in a scene with two very different levels of illumination. Dotted line is the
boundary between the two regions.
7.8Keps in 32.8ms
20Keps in 30ms
1.8Keps in 32.8ms
Fig 5.23: Some snapshots of natural scenes. From left to right: hand, woman face and man.
Figure 5.25 shows the input signals to each oscilloscope channel. Two sine wave signals
were modulated with an exponential signal. The resulting XY signal had a spiral shape. By this
way, the number of times that one individual screen point was turned ON was reduced approximately 5 times. Stimulus speed could be controlled varying the sine wave frequencies. The two
signals of Figure 5.25 were repeated periodically with a signal generator.
Figure 5.26 displays some recorded data with the Datalogger when we vary the speed of the
rotating dot. Sine wave frequencies are indicated on each graph. The average number of events is
also highlighted. The imager can detect stimuli frequencies up to 7KHz. For stimuli frequencies
higher than 1KHz, negative events (red points) are not easily detected. The reason is that negative
temporal contrast is lower than positive contrast. This can be understood with the help of Figure
5.8. In one region of the retina, we counted the number of positive and negative events generated
168
Chapter 5: The AER Temporal Contrast Vision Sensor
169
168
166
165
2.5ms
Time (ms)
167
164
163
162
100
100
50
50
0
0
x
y
Fig 5.24: Events generated by a rotating dot at 400Hz. Blue points are positive events. Red
points are negative ones. The dot stimulus was white and its background black.
by the pixels of such region during time steps of 1ms. We can notice that positive events (transition from dark to bright illumination) have a different response time than negative ones (transition
from bright to dark illumination). In fact, pixels are generating negative events all the time if the
stimulus frequency is higher than 500Hz. This suggests, that phosphor light is never completely
OFF if the stimulus is very fast. In Figure 5.27, frequency was 500Hz. Obviously, when the frequency is increased, it is more difficult to detect negative events because the decay time response
is relatively low. Positive temporal contrast is higher and positive events are generated firstly.
We can see that our sensor was able to detect rotating dots at frequencies of 10KHz
(600,000rpm). That is something that we could expect from the results of Figure 5.16. All the pixels do not have the same bandwidth. For this reason, some pixels stop spiking for lower frequencies than another ones.
If positive and negative thresholds are very low, output event rate can be very high and the
arbiter system could not send out all the events. It is desirable to work with events rates under 1
Meps to avoid information loss. If we look at the events rates of Figure 5.26, we can notice that
when stimulus frequency is lower than pixels bandwidth, the output rate is proportional to stimu-
169
Chapter 5: The AER Temporal Contrast Vision Sensor
(a)
Channel 1
Amplitude (V)
2
T=1/f
1.5
1
0.5
0
Channel 2
Amplitude (V)
2
T=1/f
1.5
1
0.5
0
Time
(b)
Fig 5.25: (a) Input signals to each oscilloscope channel. Both of them were repeated periodically. The resulting XY signal, (b), was a bright point moving through a spiral trajectory.
We varied frequency values to find out how fast was able the sensor to detect temporal contrast.
lus frequency. For stimuli frequencies higher than the cut off frequency, event rate starts to decay.
For frequencies higher than 15KHz, only the background ON events due to reset transistor leakage were measured.
170
Chapter 5: The AER Temporal Contrast Vision Sensor
500 Hz, 1.64·105 eps
2.5 KHz, 1.83·105 eps
33
26
32
Time (ms)
Time (ms)
31
30
29
25.5
25
28
27
24.5
100
100
100
50
50
100
50
0
0
x
y
50
0
y
0
x
10 KHz, 1.44·105 eps
7.5 KHz, 4.13·105 eps
30.55
8.9
30.5
30.45
Time (ms)
Time (ms)
8.8
8.7
30.4
8.6
30.35
8.5
30.3
100
100
0
y
30.25
50
50
100
0
x
50
y
0
0
50
100
x
Fig 5.26: Recording events during a XY signal period. The stimulus was a dot moving
through a trajectory with spiral shape as is shown in Figure 5.25. Results are shown for different frequencies of the input signal. Event rate is also highlighted.
171
Chapter 5: The AER Temporal Contrast Vision Sensor
(b)
(a)
120
Positive events
Negative events
45
40
100
Events/ time interval
35
y
80
60
40
30
25
20
15
10
20
5
0
0
0
20
40
60
80
100
120
20
x
25
30
35
Time (ms)
40
45
50
Fig 5.27: (b) Number of events generated by the pixels of the region the highlighted region
in (a). Time step was 1ms. Phosphor time response is slower for negative events (bright to
dark transitions) and faster positive events (dark to bright transitions). Stimulus frequency
was 500Hz
5.4.8 Power Consumption
Chip power consumption has been characterized. Supply voltage was 3.3V. Power consumption depends on static bias conditions and output rate. For this imager, it is dominated by the static
bias conditions. The main drawback of using amplifiers is an increase of the standby consumption. In further designs, it could be reduced by using amplifiers operating in subthreshold region.
Measured standby chip consumption was approximately 108mW. This dissipated power includes
the pixel reception stage, amplifiers, AGC block and I-Pots consumption. There is also a dynamic
power consumption that depends on the output rate. This consumption includes the arbiter and the
communication interface.
Typical bias settings were V ref = 1.5V , V g = 600mV , V cas = 1.8V , V bpn1 = 2V ,
I bn = 100nA , I bldc = 100nA , I brf = 10nA and I bn 1 = 200nA . Figure 5.28 shows the chip
current consumption versus the output event rate. It can vary between 132-185mW. As can be
seen, for a standard output event rate of 1Keps, total power consumption is 142mW.
5.5 Discussion
In this chapter, a new AER temporal contrast retina has been presented. Part of the work of
this thesis was to test it. The new sensor is specially useful for high speed applications and shows
some advantages over prior designs [9]. It uses amplification after the photoreceptor stage. By
this way, we achieve more bandwidth, lower latency and better fill factor. The only drawback of
amplification is more power consumption. In the future, consumption could be reduced using
amplifiers in subthreshold region. Table 5.2 summarizes the imager main specifications. We could
think that FPN would be worse in this design due to amplifiers gain mismatch. However, FPN
experimental values are similar to the values of the sensor presented by Lichsteiner et al. in [9].
172
Chapter 5: The AER Temporal Contrast Vision Sensor
58
56
Current Consumption (mA)
54
52
50
48
46
44
42
40
2
10
3
4
10
5
10
10
Output Event Rate (eps)
6
10
7
10
Fig 5.28: Chip total current consumption as a function of output event rate. Standby chip
consumption was 108mW.
Table 5.2: Sensor Specifications.
Function
ality
Temporal Contrast to
Number
of
Events
Light to
time
restriction
NO
Latency/
Bandwidth
Dynamic
Range
FPN
3.6 µs /
8KHz
>100dB
2.8%
Fill
Factor
8%
Pixel
Size
2
( µm )
35 x 35
Fabrication
Process
Power
0.35 µm
4M 2P
132185mW
The ratio between pixel deviation and gain is similar (see comparative between different sensors
in Table 1.1).
Finally, some examples of tracking of very fast stimulus (Figure 5.24 and Figure 5.26) have
been shown.
173
CHAPTER 6
Conclusions
6.1 Previous Work
The aim of this thesis was to propose new AER imagers to detect either spatial and temporal
contrast. Our first goal was to design a functional spatial contrast retina. Prior attempts to design
such sensors were plagued with mismatch and had several drawbacks like bandwidth consumption, when there was no spatial contrast. We started our work, studying the sensor presented by
Boahen [21]-[22]. After it, we decided to implement an improved version. To overcome the mismatch problem, we decided to add a new calibration system. The idea of calibration in AER systems was successfully implemented in prior designs [14]-[27]. For our particular necessities, we
proposed a new calibration method with improved features.
The second main target was to test a new AER temporal contrast sensor designed in our
group. Its main features were high speed and very low latency response. This sensor exploits
some ideas previously presented by Lichsteiner et al. in [10] and it is based on their sensor.
6.2 Achievements
In this thesis we have proposed new concepts that could be implemented in further AER
designs. The first one is calibration. We decided to design a new calibration system specially suitable for large neuromorphic arrays. Its main advantage over prior calibration systems was the fact
that calibration degrades gracefully when we scale bias currents or illumination level changes.
The calibration system is specially useful because it only needs to be calibrated once. Calibration
accuracy does not depend on illuminance strongly. Only digital registers (one for each bit of resolution) and 3 transistors have to be added to each pixel to implement the calibration procedure.
The second main achievement is the AER spatial contrast retina. It uses the calibration
method mentioned above. The sensor has also a thresholding mechanism that allows to remove
the residual mismatch after calibration. This mechanism is also useful to remove noise or inhibit
the activity of positive or negative events. There are two independent channels to process positive
or negative events. For some applications, it could be desirable to work with unsigned events. In
these cases, the thresholding mechanism is an efficient method to inhibit completely the activity
175
Chapter 6: Conclusions
of positive or negative events. Moreover, the imager has an optional TFS (time-to-first spike)
mode. It is implemented with a global reset mechanism. By adjusting the voltage of the refractory
period, the number of spikes that one individual pixel generates can be controlled. If we set the
refractory period to allow each pixel just spike once after the global reset signal becomes active,
pixels with more activity (more relevant information) will spike first. Thus, if the reset period is
chosen suitably, only relevant information will be transmitted. We can achieve a frame-based processing with some advantages of AER systems. Another important features of the sensor are high
sensitivity, low fixed pattern noise and low latency.
The second proposed sensor was a temporal contrast retina. Its main features were high
speed, low latency and high dynamic range. This sensor has a similar topology to the one proposed by Lichsteiner et al. [10]. Its main advantages are the use of amplifiers after the photoreception stage. By this way, temporal contrast sensitivity and speed response can be improved. At the
same time, fill factor was reduced because the gain at the capacitance divider of the differencing
circuit could be decreased.
Extensively experimental results showing the validity of the new calibration systems and the
main features of the two contrast retinae are provided. In chapter 2 a new circuit to generate biasing currents was presented. In chapter 3, the calibration circuitry is explained, in chapter 4 the
spatial contrast retina is presented, and in chapter 5, the temporal contrast retina is showed.
6.3 Future Work
After studying the new circuits thoroughly, some improvements could be proposed for each
one. Let us start with the calibration system.
As we explained in chapter 3, the calibration procedure is presently off-line, but it is perfectly possible to implement it on-chip with an extra controller (designed and synthesized via
VHDL, for example), since it only requires to expose the chip to uniform illumination (one can
simply remove the optics), compare the pixel frequencies (we would only need to count the number of times that each pixel spikes during a time interval) and compute an optimum set of calibration weights (find the calibration words that minimize the number of spikes of each pixel).
If we speak about the spatial contrast retina, the feasibility of the design has been shown.
However, the retina resolution is rather low (32 x 32). It would be easy to design a new prototype
with more resolution to take images with more quality. One of its drawbacks is the low fill factor.
This could be easily improved in the future. The FPN of the circuit is very low (0.6%), so we
could reduce the number of bits (3 or 4 bits registers would be enough to calibrate the chip) of the
calibration system keeping a reasonable FPN value. Combining calibration and thresholding, better results can be achieved. Maybe, a more precise thresholding mechanism could be designed. It
would only be necessary to increase the transistor size of the two transistors of the current mirror
used to implement the thresholding method. Moreover, an extra controller to monitor the time-tofirst spike mode could be added. It could control the duration of the frames, Tframe, or the number
of spikes received before resetting the sensor.
With regard to the AER temporal contrast retina, we improved the sensor speed and
response time with an amplification stage. We also increased fill factor at the same time. However, amplifiers operate in strong inversion and their power consumption is higher in comparison
to the rest of blocks that make up the pixel. In further designs, amplifiers could be designed to
work in weak inversion and reduce its power consumption. Pixel size is relatively low
2
( 35x35µm ). It could be reasonable to implement a temporal contrast retina with more resolu-
176
Chapter 6: Conclusions
tion. An optional time-to-first spike mode could be added to this sensor. Only 3 extra transistors
per pixel and one output pin would be necessary to have available the optional operation mode.
177
APPENDIX A
Abbreviations and Mathematical Conventions
Throughout this thesis, we used the following abbreviations,
CCD
Charge-Coupled Device
VLSI
Very Large Scale Integration
Matlab
Mathematical Program of Mathworks, Inc
APS
Active Pixel Sensor
FPN
Fixed Pattern Noise
Iph
Photo current
V
Voltage
AER
Address Event Representation
VHDL
Very High Speed Integrated Circuit Hardware Description Language
The following mathematical symbols are also used,
f, w
Temporal frequency
T
Temporal period
Absolute value
179
Appendix A
w cal
Optimum calibration word
θ
Temporal contrast
∆V diff_ev
Voltage threshold to generate one event (either positive or negative)
∆V diff
Voltage variation in response to temporal contrast stimulus
θ ev
Contrast threshold to generate one event (either positive o negative)
σ θ ev
Deviation of the contrast threshold
θ min
Minimum detectable contrast.
Gi
Gain of the amplification stage, i
τ
System response time
180
Bibliography
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
M. Sivilotti, “Wiring considerations in analog VLSI systems with application to field-programmable networks,” Ph.D. dissertation, Calif. Inst.Technol., Pasadena, 1991.
M. Mahowald, “VLSI analogs of neural visual processing: a synthesis of form and function” Ph.D. dissertation, California Institute of Technology, Pasadena, 1991.
T. Delbrück, Library essentials, Analog VLSI and Neural Systems by Carver Mead, Addison Wesley, 1986. The Neuromorphic Engineer, 1(1):11, 2004. http://ine-web.org/research/
newsletters/index.html.
Kwabena A. Boahen, “Retinomorphic Vision Systems: Reverse Engineering the Vertebrate
Retina”, Ph. D. dissertation, California Institute of Technology, Pasadena, 1996.
G. M. Shepherd, The Synaptic Organization of the Brain, 3rd ed. Oxford, U.K.: Oxford University Press, 1990.
M. Mahowald. An Analog VLSI System for Stereoscopic Vision, Kluwer, Boston, MA, 1994.
U. Mallik, et al., “Temporal Change Threshold Detection Imager”, in ISCCC Dig. of Tech.
Papers, San Francisco, 2005, pp362-363.
Y. Chi et al., “CMOS Camera with In-pixel Temporal Change Detection and ADC”, IEEE
Journal of Solid State Circuits, vol. 42, pp, 2187-2196, OCT 2007.
Patrick Lichsteiner et al, “A 128x128 120dB 15 µs Latency Asynchronous Temporal Contrast Vision Sensor”, IEEE Journal of Solid State Circuits, vol. 43, pp. 566-576, 2008.
P. Lichsteiner, et al., “A 128x128 120dB 30mW Asynchronous Vision Sensor that Responds
to Relative Intensity Change”, in ISSCC Dig. of Tech. Papers, San Francisco, 2006, pp.
508-509 (27.9)
S. Thorpe, D. Fize, C. Marlot, “Speed of Processing in the Human Visual System”, Nature
381:520-2, 1996.
X. G. Qi, X.; Harris J., “A Time-to-first-spike CMOS imager”, in Proc. of the 2004 IEEE
International Symposium on Circuits and Systems (ISCAS04), Vancouver, Canada, 2004,
pp. 824-827.
J. A. Leñero-Bardallo, T. Serrano-Gotarredona, and B. Linares-Barranco, “A Calibration
Technique for Very Low Current and Compact Tunable Neuromorphic Cells. Application to
181
BIBLIOGRAPHY
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
5-bit 20nA DACs”, IEEE Transactions on Circuits and Systems, Part-II: Brief Papers,
vol.55, No. 6, pp. 522-526, June 2008.
J. Costas-Santos, T. Serrano-Gotarredona, R. Serrano-Gotarredona and B. Linares-Barranco, “A Spatial Contrast Retina with On-chip Calibration for Neuromorphic Spike-Based
AER Vision Systems”, IEEE Trans. Circuits and Systems, Part-I: Regular Papers, vol. 54,
No. 7, pp. 1444-1458, July 2007.
M. Barbaro, P.-Y. Burgi, A. Mortara, P. Nussbaum, and F. Heitger. “A 100x100 Pixel Silicon
Retina for Gradient Extraction with Steering Filter Capabilities and Temporal Output Coding”. IEEE Journal of Solid-State Circuits, 37(2):160-172, Feb. 2002.
P. F. Ruedi, et al., “A 128x128 Pixel 120-dB Dynamic-range Vision-Sensor Chip for Image
Contrast and Orientation Extraction”, IEEE Journal of Solid-State Circuits, 38:2325-2333,
Dec. 2003.
P. F. Ruedi, et al., “An SoC Combining a 132dB QVGA Pixel Array and a 32b DSP/MCU
Processor for Vision Applications”, in IEEE ISSCC Dig. of Tech. Papers, 2009, pp.46-47,
47a.
Honghao Ji and Pamela A. Abshire, “Fundamentals of Silicon-Based Phototransduction” in
“CMOS Imagers from Phototransduction to Image Processing”, Kluwer Academic Publishers.
E. Culurciello, R. Etienne-Cummings, and K. A. Boahen, “A Biomorphic Digital Image
Sensor”, IEEE J. Solid-State Circuits, vol.38, pp.281-294, 2003.
S. Chen, and A. Bermak, “Arbitrated Time-to-First Spike CMOS Image Sensor with OnChip Histogram Equalization”, IEEE Transactions VLSI Systems, vol. 15, No. 3, pp 346357, March 2007.
K. A. Zaghloul and K. Boahen, “Optic Nerve Signals in a Neuromorphic Chip: Part 1”,
IEEE Transactions on Biomedical Engineering, vol 51, pp. 657-666, 2004.
K. A. Zaghloul and K. Boahen, “Optic Nerve Signals in a Neuromorphic Chip: Part 2”,
IEEE Transactions on Biomedical Engineering, vol 51, pp. 667-675, 2004.
J. A. Leñero-Bardallo, T. Serrano-Gotarredona, and B. Linares-Barranco, “A 5-Decade
Dynamic Range Ambient-Light-Independent Calibrated Signed-Spatial-Contrast AER Retina with 0.1ms Latency and Optional Time-to-First-Spike Mode”, Transactions on Circuits
and Systems, Part I, Under press.
C. Posch, et al., “High DR, Low Date-rate Imaging Based on an Asynchronous, Self-triggered Address-event PWM Array with Pixel-level Temporal Redundancy Suppression”, in
2010, in ISCAS 2010.
R. R. Harrison, J.A. Bragg, P. Hasler, B.A. Minch, and S.P. DeWeerth, “A CMOS
programmable analog memory-cell array using floating-gate circuits,” IEEE Trans. on
182
BIBLIOGRAPHY
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
Circuits and Systems, Part II, vol. 48, No. 1, pp. 4-11, Jan. 2001.
Y. L. Wong, M. H. Cohen, and P. A. Abshire, “128x128 floating gate imager with selfadapting fixed pattern noise reduction,” Proc. of the IEEE 2005 Int. Symp. on Circuits and
Systems (ISCAS’05), vol. 5, pp. 5314-5317, 2005.
R. Serrano-Gotarredona, T. Serrano-Gotarredona, A. Acosta-Jimenez, and B. LinaresBarranco, “A Neuromorphic Cortical-Layer Microchip for Spike-Based Event Processing
Vision Systems,” IEEE Trans. on Circuits and Systems I: Regular Papers, vol. 53, No. 12,
pp. 2548-2566, Dec. 2006.
B. Linares-Barranco, T. Serrano-Gotarredona, and R. Serrano-Gotarredona, "Compact LowPower Calibration Mini-DACs for Neural Massive Arrays with Programmable Weights,"
IEEE Trans. on Neural Networks, vol. 14, No. 5, pp. 1207-1216, September 2003.
C. A. Laber, C. F. Rahim, S. F. Dreyer, G. T. Uehara, P. T. Kwok, and P. R. Gray, “Design
Considerations for a high-Performance 3mm CMOS Analog Standard-Cell Library,” IEEE
J. Solid-State Circuits, vol. SC-22, No. 2, pp. 181-189, April 1987.
T. Delbrück and A. Van Shaik, “Bias Current Generators with Wide Dynamic Range” Int.
Journal of Analog Integrated Circuits and Signal Processing, No. 43, pp. 247-268, June
2005.
T. Delbrück and A. Van Shaik, “Bias Current Generators with Wide Dynamic Range” ,
2004 International Symposium on Circuits and Systems (ISCAS 2004), Vancouver, Canada,
May 23-25 2004 pp. I-337-340.
T. Delbruck, P. Lichtsteiner, R. Berner, C. Dualibe, “32-bit Configurable bias current generator with sub-off-current capability”, ISCAS 2010, (in press)
R. R. Harrison, J.A. Bragg, P. Hasler, B.A. Minch, and S.P. DeWeerth, “A CMOS
programmable analog memory-cell array using floating-gate circuits,” IEEE Trans. on
Circuits and Systems, Part II, vol. 48, No. 1, pp. 4-11, Jan. 2001.
K. Bult and G.J.G.M. Geelen, “An inherently linear and compact MOST-only current
division technique,” IEEE J. Solid-State Circuits, vol. 27, No. 12, pp. 1730-1735, Dec.
1992.
T. Serrano-Gotarredona and B. Linares-Barranco, "CMOS Mismatch Model valid from
Weak to Strong Inversion", Proc. of the 2003 European Solid State Circuits Conference,
(ESSCIRC’03), pp. 627-630, September 2003.
P.R. Gray, P.J. Hurst, S.H. Lewis, and R. G. Meyer, Analysis and Design of Analog
Integrated Circuits, 4th Edition, John Wiley, 2001.
Rafael Serrano-Gotarredona, Luis Camuñas-Mesa, Teresa Serrano-Gotarredona, Juan A.
Leñero-Bardallo and Bernabé Linares-Barranco, “The Stochastic I-Pot: A Circuit Block for
Programming Bias Currents”, IEEE Transaction on Circuits and Systems-II: Brief Papers,
183
BIBLIOGRAPHY
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
vol 19, No. 7, pp. 1196-1219, July 2008.
B. Linares-Barranco and T. Serrano-Gotarredona, “On the Design and Characterization of
Femtoampere Current-Mode Circuits”, IEEE Journal of Solid-State Circuits, vol. 38, No. 8,
pp. 1353-1363, August 2003.
M. Azadmehr, J. Abrahamsen, and P. Häfliger, “A Foveated AER Imager Chip”, in Proc. of
the IEEE International Symposium on Circuits and Systems (ISCAS2005), pp. 2751-2754,
Kobe, Japan, 2005.
R.J. Vogelstein, U. Mallik, E. Culurciello, R. Etienne-Cummings, and G. Cauwenberghs,
“Spatial acuity modulation of an address-event imager,” Proc. of the 2004 11th IEEE
International Conference on Electronics, Circuits and Systems (ICECS 2004), pp. 207-210,
Dec. 2004.
J. Kramer, “An Integrated Optical Transient Sensor”, IEEE Transactions on Circuits and
Systems, Part II: Analog and Digital Signal Processing, v. 49, No. 9, pp. 612-628, Sep.
2002.
M. Arias-Estrada, D. Poussart, and M. Tremblay, “Motion Vision Sensor Architecture with
Asynchronous Self-Signaling Pixels”, Proceedings of the 7th International Workshop on
Computer Architecture for Machine Perception (CAMP07), pp. 75-83, 1997.
K. Boahen, “Retinomorphic Chips that see Quadruple Images”, Proceedings of International Conference of Microelectronics for Neural, Fuzzy and Bio-inspired Systems
(Microneuro99), pp. 12-20, Granada, Spain, 1999.
S. Thorpe, et al., “SpikeNet: Real-time Visual Processing in the Human Visual System”,
Neurocomputing 58-60: 857-64, 2004.
K. Boahen and A. Andreou, “A Contrast-Sensitive Retina with Reciprocal Synapses”, in J.
E. Moddy (Ed.), Advances in neural information processing, vol. 4, pp- 764-772, San
Mateo CA, 1992. Morgan Kaufman.
A. G. Andreou and K. Boahen, “Translinear Circuits in Subthreshold CMOS”, Analog Integrated Circuits and Signal Processing, Kluwer, no. 9 , pp. 141-166, Apr. 1996.
K. Boahen, “Point-to-Point Connectivity Between Neuromorphic Chips Using Address
Events”, IEEE Transactions on Circuits and Systems Part-II, vol. 47, No. 5, pp. 416-434,
May 2000.
R. Berner, T. Delbruck, A. Civit-Balcells and A. Linares-Barranco,“A 5 Meps $100 USB2.0
Address-Event Monitor-Sequencer Interface”, IEEE International Symposium on Circuits
and Systems,2007, ISCAS 2007, pp. 2451-2454.
jAER Open Source Project 2007 [Online]. Available: http://jaer.wiki.sourceforge.net.
V. Gruev and Etienne-Cummings, R. A. pipelined temporal difference imager. IEEE J. Solid
State Circuits, 39:538-543, 2004.
184
BIBLIOGRAPHY
[51] M. A. Mahowald, “VLSI Analogs of Neuronal Visual Processing: A Synthesis of Form and
Function”, PhD, Computation and Neural Systems, Caltech, Pasadena, California, 1992.
[52] D. A. Baylor, T. D. Lamb, and K.W. Yau, “Responses of Retinal Rods to Single Photons”, J.
Physiology., 288:613-634, 1979.
[53] D. A. Baylor, B. J. Nunn, and J. L. Schnapf. “The Photo Current, Noise, and Spectral Sensitivity of Rods of the Monkey Macaca Fascicularis”. J. Physiology, 357:575-607, 1984.
[54] Patrick Lichtsteiner, “An AER Temporal Contrast Vision Sensor”, Ph.D. Dissertation, ETH
Zürich, 2006.
[55] Foveon , Inc., Website, www.foveon.com.
[56] H.R. Blackwell. Contrast threshold of the human eye, “J. Opt. Soc. Am., 36:624-643, 1946.
[57] P. Sterling, E. Cohen, M. Freed, and R.G. Smith. “Microcircuitry of the On-beta Ganglion
Cell in Daylight, Twilight, and Starlight”, Neuroscience. Res. (Suppl.), 6:269-285,1987.
[58] D.C. Burr. “Motion Smear”, Nature, 284:164-165, 1990.
[59] Tobi Delbrück, Bernabe Linares-Barranco, Eugenio Culurciello, Christoph Posch, “Activity-Driven, Event-Based Vision Sensors”, IEEE International Symposium on Circuits and
Systems, 2010, ISCAS 2010, Paris.
[60] J. Zihl, D. Von Cramon, and N. Mai. “Selective Disturbance of Movement Vision After
Bilateral Brain Damage”. Brain, 106(2):313-340.
[61] J. Kramer, “An ON/OFF Transient Imager with Event-driven, Asynchronous Readout”, In
Procc. IEEE International Symposium on Circuits and Systems, May 2002.
[62] J. Kramer, “An Integrated Optical Transient Sensor”, IEEE Transactions of Circuits and
Systems II, 49(9):612-628, Sep 2002.
[63] T. Delbrück and C.A. Mead. “Adaptive Photoreceptor with Wide Dynamic Range”, In 1994
IEEE International Symposium On Circuits and Systems”, volume 4, pages 339-342, London, May 1994. ISCAS’94: London, England, 30 May-2 June.
[64] J. Kramer, R. Sarpeshkar, and C. Koch. “Pulse-based Analog VLSI Velocity Sensors”, IEEE
Transactions of Circuits and Systems II, 44(2):86-101, February 1997.
185
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement