Driver-Vehicle Interaction Identification, Characterization and Modelling of Path Tracking Skill Andreas Erséus

Driver-Vehicle Interaction Identification, Characterization and Modelling of Path Tracking Skill Andreas Erséus
Driver-Vehicle Interaction
Identification, Characterization and
Modelling of Path Tracking Skill
Andreas Erséus
Doctoral Thesis in Vehicle Engineering
TRITA-AVE 2010:29
ISSN 1651-7660
ISBN 978-91-7415-665-2
Postal address
Royal Institute of Technology
KTH Vehicle Dynamics
SE-100 44 Stockholm, Sweden
Visiting address
Teknikringen 8
Stockholm
Telephone
+46 8 790 6000
Telefax
Internet
+46 8 790 9290 www.ave.kth.se
© Andreas Erséus 2010
Abstract
Since the dawn of the automobile, driver behaviour has been an issue. Driving can
result in accidents that may harm not only the driver but also passengers and the
surroundings. This calls for measures that restrict the usage of vehicles and to assist the
individual driver to conduct the driving in a safe, yet practically efficient manner. The
vehicles should therefore be both safe and intuitive, and preferably answer to the
different needs of all kinds of drivers.
Driving skill can be defined in many ways, depending on the objective of the driving
task, but answer in some way to the question of how well the driver can conduct the
driving task. To assist low skill drivers without compromising the driving demand for
high skill drivers, it is of highest importance that vehicles are tested and designed to
meet those needs. This includes both the testing activities in the vehicle design phase in
general but also the configuration for active systems and preventive safety, preferable
with settings that adapts to the skill of the individual driver.
The work here comprises the definition of skill and of driver recruitment procedures,
scenario design, the development of an analysis method for objective measures, and the
gathering of metrics to characterize the driver skill. Moreover, a driver model has been
developed that makes use of driver skill characteristics. To gather the information
needed, extensive multidisciplinary literature studies were conducted, as well as using
field tests and test using an advanced moving base driving simulator. Here the focus is
on path tracking skill, which is the main control aspect of driving, although the
developed driving scenarios allow a varying degree of path planning, which is more
related to regulation. The first simulator test was done with a very simple criterion for
driver selection, but the results gave a good insight into the variation between drivers in
general. For the following tests the recruitment procedure was refined to find drivers
with high or low vehicle control and regulation skill, a recruitment that also was
verified to really represent two different populations.
A method was defined that successfully identified sets of skill-related measures, with
some variation in composition depending on the path tracking demand on the driver. In
the curving road scenario, for example, the highest number of skill-related measures is
identified in the curves, which is reasonable since the straight segments do not require
the same amount of active control from the drivers.
The driver model developed uses a quasi-static analytical description of the driver
knowledge of the vehicle dynamics, but possesses the capability of nonlinear
descriptions. The parameters in this model are mainly physical properties that easily can
be related to the driving process. Metrics gathered are used for identification of the
driver model setup for a double lane change scenario using an optimization routine,
with adjusted parameter settings for different velocities.
With a subjective comparison of the recorded driving simulator data, the method is
verified to enable driver skill settings for driver models. In addition, the method allows
metrics to be gathered for driver skill identification routines, meeting the defined
objectives of the project.
i
ii
Preface
Can one vehicle be designed to meet every possible demand? Well, that might be
difficult to achieve, but a regular car should at least be driveable by a wide range of
different people with driving licences issued under different conditions. Understanding
the behaviour of these people might be one of the biggest challenges in vehicle design
today. The latest advancements in preventive safety used for focused driver support
could be a great aid, especially for less skilled drivers, but this also introduces a risk of
interfering with the drivers’ intentions. What really inspired me to take on the challenge
of doing this thesis work were the possibilities of active vehicle adjustments to best suit
the current driver and state that could emerge if the specific driver’s needs is known.
This leads us to the intriguing question that appeared in the beginning of this work: Is it
really feasible to measure and model driving skill using in-vehicle sensors?
Acknowledgements
This research project has been performed at KTH Vehicle Dynamics, in collaboration
with General Motors North America, Saab Automobile and VTI (the Swedish National
Road and Transport Research Institute). The financial support by VINNOVA (The
Swedish Agency for Innovation Systems) is also gratefully acknowledged.
I would like to express my gratitude to my academic supervisor, Professor Annika
Stensson Trigell, for the support provided during the project, with her unceasing
positive spirit and encouragement. Thanks are also extended to Lars Drugge, who has
been a co-advisor for me during the later part of the project, with fruitful discussions
and straightforward feedback. Special thanks are given to Staffan Nordmark, also a coadvisor, who put me on the right track after some setbacks and devoted many hours to
provide valuable input in the development of the driver model. Thanks are also due to
Gunnar Olsson for his feedback and hearty interest in the project, inspiring us with his
long experience of vehicle design and testing, to Arne Nåbo for providing much
appreciated expertise in human factors and Steve Chin for encouraging us to strive for
more knowledge. I am also grateful for the fantastic support given by Håkan
Sehammar, Göran Palmkvist and Mats Lidström at VTI. Present and past colleagues,
first and foremost Markus Agebro, Johan Andreasson and Jonas Jarlmark that were at
KTH when I started, it has been a real privilege to work with you during these years.
My wonderful newly wedded wife Emmeli, you have been unbelievably supporting
despite the long hours which I had to devote to this work. My parents, relatives and
friends, thank you all for encouraging me and for being there for me when I needed.
Even though some of you are not with us any more, I know you never doubted me to
complete this work. You have been the best sources of inspiration and still are.
Andreas Erséus
Stockholm, Sweden, May 2010
iii
iv
Appended papers
Paper A
Nilsson, A., Agebro, M. and Stensson Trigell, A. Study of path tracking skill and
strategy using a moving base simulator, FISITA Transactions, paper F2006D075T,
2007.
Contributions of authors: Erséus (formerly Nilsson) and Agebro designed the
scenario, recruited the drivers and supervised the test. Agebro implemented the vehicle
model’s steering servo. Erséus did the analysis and wrote the paper. Stensson Trigell
and Drugge provided useful ideas, valuable comments and proofread the paper. Erséus
also presented the paper at FISITA’06 World Automotive Congress, Yokohama, Japan,
October 22-27, 2006.
Paper B
Erséus, A., Stensson Trigell, A. and Drugge, L. Methodology for finding parameters
related to path tracking skill applied on a DLC-test in a moving base driving simulator,
submitted for publication (this is an extended version of Nilsson, A., Stensson Trigell,
A. and Drugge, L. Methodology to find parameters characteristic to path tracking skill
– DLC-test in a moving base simulator, Proceedings of the 21st International
Symposium on Dynamics of Vehicles on Roads and Tracks (IAVSD’09), Stockholm,
Sweden, 2009).
Contributions of authors: Erséus (formerly Nilsson) developed the recruitment criteria
and recruited the drivers (with assistance from Markus Agebro). Erséus developed the
analysis method, did the analysis and wrote the paper. Stensson Trigell and Drugge
provided useful ideas, valuable comments and proofread the paper. Erséus also
presented the original paper at IAVSD’09.
Paper C
Nilsson, A., Stensson Trigell, A. and Drugge, L. A path tracking scenario without
preview for analysis of driver characteristics, Proceedings of the 9th International
Symposium on Advanced Vehicle Control (AVEC’08), Kobe, Japan, 2008.
Contributions of authors: Erséus (formerly Nilsson) designed the scenario, did the
analysis and wrote the paper. Stensson Trigell and Drugge provided useful ideas,
valuable comments and proofread the paper. Erséus also presented the paper at
AVEC’08.
v
Paper D
Erséus, A. Drugge, L. and Stensson Trigell, A. A path tracking driver model with
representation of driving skill, submitted for publication.
Contributions of authors: Erséus (formerly Nilsson) designed the model (with
assistance from Staffan Nordmark), did the analysis and wrote the paper. Stensson
Trigell and Drugge provided useful ideas, valuable comments and proofread the paper.
Paper E
Erséus, A. Stensson Trigell, A. and Drugge, L. Characteristics of path tracking skill on
a curving road, submitted for publication.
Contributions of authors: Erséus (formerly Nilsson) designed the scenario, did the
analysis and wrote the paper. Stensson Trigell and Drugge provided useful ideas,
valuable comments and proofread the paper.
vi
Contents
1 Introduction............................................................................................................1
1.1
1.2
1.3
1.4
1.5
1.6
History of driving and introduction to driving skill ......................................... 1
Preventive safety and driver adaptation ........................................................ 2
Path planning and tracking ............................................................................. 3
Driver models ................................................................................................. 4
Objectives ....................................................................................................... 5
Outline of thesis.............................................................................................. 5
2 Method...................................................................................................................7
2.1
2.2
Work process .................................................................................................. 7
Limitations ...................................................................................................... 9
3 The human vehicle operator .................................................................................11
3.1
3.2
3.3
General driver characteristics....................................................................... 11
The driving process ....................................................................................... 13
3.2.1
Driving functions and tasks............................................................. 13
3.2.2
Specification of driving activities .................................................... 19
Individual experience-based differences...................................................... 19
3.3.1
Behaviour development ................................................................. 19
3.3.2
Driving skill and driver performance .............................................. 23
3.3.3
Driving style .................................................................................... 25
3.3.4
Specification of ordinary driving and path tracking skill................. 28
4 Measurement of driver characteristics .................................................................29
4.1
4.2
4.3
4.4
4.5
Introduction to driver analysis...................................................................... 29
Examples of measurements.......................................................................... 30
Driving simulator tests.................................................................................. 33
4.3.1
Test platform .................................................................................. 34
4.3.2
Curved cone track scenario ............................................................ 35
4.3.3
Avoidance manoeuvre scenario ..................................................... 36
4.3.4
Driver response scenario ................................................................ 36
4.3.5
Curving road scenario ..................................................................... 37
Proposed driving skill characterization methodology .................................. 38
Results from driving characterization........................................................... 38
5 Modelling of vehicles ............................................................................................45
5.1
5.2
5.3
MBS-model description ................................................................................ 45
Validation of MBS-model.............................................................................. 46
Driver’s internal vehicle model..................................................................... 47
vii
6 Modelling of drivers..............................................................................................51
6.1
6.2
6.3
6.4
Compensation tracking models .................................................................... 51
Preview tracking models............................................................................... 52
Fuzzy set theory models ............................................................................... 56
The KTH Vehicle Dynamics driver model ...................................................... 59
7 Scientific contributions .........................................................................................61
8 Discussion and conclusions ...................................................................................63
9 Recommendations for future work.......................................................................67
References................................................................................................................69
Nomenclature ..........................................................................................................73
viii
Chapter 1
Introduction
This thesis begins with a background of the problem and relevance for regular drivers.
Additionally, the opportunities for the future, safety concerns and the potential to
enhance the driving experience for all drivers are presented, followed by the defined
objectives and outline of thesis.
1.1 History of driving and introduction to driving skill
Since the day when the first man-operated motorized road vehicle appeared, driver
behaviour and driving skill have been an issue. In Britain, with the Locomotive Act of
1861, the speed of all horseless vehicles was limited to 10 mph outside towns and 4
mph within, which was mainly due to the steam alarming horses and the vehicles
harming the roadways. In the Locomotive Act of 1865, not only was the speed limit
lowered to only 4 mph and 2 mph for driving outside and inside towns respectively, but
also it was required that a man with a red flag or a lantern should walk 60 yards ahead
of each vehicle in towns to enforce a walking pace and to warn about this self-propelled
machine [1,2]. Speed regulation is one severe way of limiting the potential danger for
the driver and the surroundings used for most roads today, since both the risk of an
accident and the result of such an event can be affected by the vehicle speed. Controlled
driver support systems are also becoming increasingly important in their role of helping
drivers to perform safe driving. However, these and other safety measures, both
regulatory and technical, often rely on an average of driver performance and rarely take
individual drivers into consideration. This may place unnecessarily hard and sometimes
counterproductive restrictions on high skill drivers, while low skill drivers could benefit
from more severe interference. Moreover, even when they are not specifically
interfering, different setups may prove to be best in supporting different driver types.
Skill is a very broad term that can include many different components, but it usually is
a measure on how effective a pre-defined task is completed. The task can be defined at
different levels, which for driving can be complex tasks as for example winning races
or to go from point A to B without any incidents. It can also be as small as shifting the
gear at an appropriate time or spotting wild life at the side of the road. For all these
1
Chapter 1 Introduction
2
tasks a person evolves a process of executing the task in a more coordinated or
automated way with less mental workload involved and/or less error, which can be
considered as becoming more skilled. Specific limited tasks, such as grasping and
turning knobs, have been thoroughly investigated in many papers, e.g. in [3], where
strategy in knob turning was studied (e.g. arm motion and applied torque), and in [4],
where human properties for grasping a knob were identified. For very simple tasks the
individual human physical properties can have a large effect, and the amount of training
can be of minor importance. These and other aspects of human limitations and learned
behaviour will be further explored later in this thesis, see Chapter 3.
1.2 Preventive safety and driver adaptation
Driving consists of highly complex tasks and it would seem preferable if the driver
were to use his full potential on the aspects of driving that require decision-making and
intelligent thinking. To make that possible, system can be introduced that reduce
information uncertainty, correct errors and adapt the vehicle to the individual driver and
the situation. Countermeasures that could enhance information acquisition and
responses are listed in [5], with special emphasis on operational systems which guide
response and direct attention and could support elderly drivers, as well as systems
which act on a strategic or tactical level and could be of great benefit to young people,
compensating for their inexperience (using a general average of age-typical problems).
Piechulla et al present in [6] a situation-adaptive man-machine interface for the
optimum allocation of drivers’ attention resources, where they use a demonstrator
vehicle with a developed version of a “state-of-the-art” (as it was at that time) adaptive
cruise control system (ACC) based on radar sensors and an experimental heading
control system (HC) based on computer vision. The HC in this system searches for lane
markings and employs small forces on the steering wheel which serve as indicators of
how to steer in order to stay in the lane. There is a range of sensors available for
probing the environment around the vehicle, and these sensors can be used both as a
source of information to the driver and as input to vehicle systems (Figure 1); but the
sensors can also be used for determining both the vehicle’s and the driver’s internal
status (directly or indirectly from the coupled system).
Video
Long Range
Radar
Figure 1:
Infrared
Video
Medium
Ultralong
Night vision
range
2 ≤ 150 m
≤ 200 m
0 ≤ 80 m
Short Range
(Radar, Lidar)
Short
0.2 ≤ 20 m
Ultrasonic
Ultrashort
0.2 ≤ 1.5
(2.5) m
Range of some available sensors for analyzing vehicle surroundings [7].
Section 1.3 Path planning and tracking
Brännström highlights the importance of a good balance between not disturbing the
driver in real driving conditions, and supporting the driver in critical situations [8].
With more knowledge of the expected behaviour of drivers, which can vary greatly
between individuals, a more elaborate decision-making is possible for the engagement
of a preventive safety system in the vehicle. Stability systems that make use of the
adjustability of the controlled system, e.g. the adjustability of the steering wheel rate
and effort, or the suspension characteristics, might also benefit from knowledge of the
effect of different settings on different driver types. The driving workload and
distraction aspects could also possibly be improved, e.g. by only showing relevant
information when a driver is driving close to the maximum of his or her ability.
1.3 Path planning and tracking
Driving requires many different processes to be performed, and path planning and path
tracking are terms that are used for two of the elements.
Path planning is here referred to as the driver’s mental process of determining the path
along which he or she wants the vehicle to go (in a short timeframe, i.e. a typical
duration of less than 1 minute, as part of the regulating loop [9]). This path can be more
or less given by the road boundaries, but usually there is enough space to allow some
variation in the driver’s choice of the path to track. Moreover, when there is a
possibility of selecting action or timing, for example passing a vehicle or performing a
lane change, the driver needs to plan the procedure. Since the vehicle performance and
driver capabilities can differ, good planning requires some knowledge of what can be
accomplished (with the system capabilities) without causing an accident.
Path tracking is a term used for the process of the driver following the chosen path (i.e.
a typical duration of not more than 1 s, as part of the tracking loop [9]). This can be
carried out with greater or lesser accuracy. Both the vehicle and the driver impose delay
in the system when a difference between the preferred path and the vehicle’s path is
detected. Basic physical effects that slow down the system response are for the vehicle
the inertia and tyre force build-up, for example, and for the driver the cognitive process
and muscle actuation, for example. This means that good path tracking requires a feedforward process that compensates for the delays. To enable prediction of the required
action, the drivers need to have some level of knowledge of how the vehicle will react
on input. Krendel and McRuer divided the tracking characteristics into three basic
levels in 1960: the compensatory level, i.e. compensation for observed errors; the
pursuit level, which takes experience and prediction into account, although operating in
a closed loop with visual feedback, for example; and the precognitive level, which is
essentially open loop control based on knowledge of the system and the required future
inputs [10]. The third level is not directly considered to be tracking since it is open
loop, but it is indirectly influencing the tracking since it is conducted to minimize future
path errors.
3
Chapter 1 Introduction
4
1.4 Driver models
In many situations it can be beneficial to do analysis of the driver and vehicle with one
or the other, or both, replaced by a virtual representation. For real vehicle tests you can
replace the driver with a steering robot, being superior to the human in precision and
repeatability for pre-defined control of the vehicle. If you instead want to do tests
without a physical vehicle but a real driver you can use a driving simulator with a
model of the vehicle implemented. This provides a controlled and safe environment,
and allows you to do tests with high measurement accuracy that could be dangerous or
even impossible to do in reality. Many different vehicle configurations and
environments can be tested in a single driving session, which can reduce the number of
vehicle prototypes that actually has to be built. If both the vehicle and the driver are
replaced with computer models, this allows you to do large batches of tests in desktop
computer simulation programs. For the driver input to the vehicle model you can use
either open loop pre-defined steering (similar to the steering robot input) or driver
models, which more accurately should represent the human driver with his or her
limitations.
One of the first recognised model based driver descriptions, prior to any desktop
computer system, is to be found in [11] from 1938 by Gibson and Crooks. It treats the
vehicle and obstacles as repelling forces that become stronger when the risk of an
accident is higher, shaping a “field of safe travel” (Figure 2). A rise in the number of
papers on driver models began in the late fifties and the production of papers continued
at a high level into the eighties. McRuer is one author who has had great influence on
control-theory-based pilot models and driver model development, e.g. in [12,13,14],
and with Wier as one co-author in many papers with more emphasis on drivers, e.g.
[15,16,17,18,19,20]. Other authors who have pioneered the development of driver
models are, to name a few, Fiala (e.g. [21]), Mitschke (e.g. [22]), Allen (e.g. [19,23]),
and Donges [24]. In the early eighties, MacAdam presented his work on optimal control
[25,26], which provided a much appreciated method for predicting vehicle movement
which allowed good path following. In [27] this is put in a larger context with driver
limitations included. Sharp has among other things contributed with mathematical
model [28] and optimal control model [29] development. Another driver modelling
approach was given by Cole, who has studied neuro-muscular activities in the driver’s
steer control and implemented this research in driver models, e.g. in [30] by Pick and
Cole.
Figure 2:
Example showing the field of safe travel in an intersection [10].
Section 1.5 Objectives
Driver models are used not only for simulation of the individual driver, which is the
main usage of the driver models referred to here, but also to account for interaction
between drivers. These models can be simplified for low-level control, but have to be
more elaborate for other levels. In the thesis on driver modelling for transport systems
by Ma [31], the driver tasks are divided into three levels, namely the strategic, tactical
and operational levels, which have corresponding level descriptions within the field of
driver modelling on the individual level, e.g. navigation, guidance and control. The
difference is that the focus of driver modelling for traffic simulation is mainly directed
on the second level, the tactical [31], while the main focus for individual driver
modelling is directed on the operational level. In recent years, however, patternrecognition, using neural networks, for example, has gained increased attention in
studies of the tactical behaviour of individual drivers, e.g. the study by
Raksincharoensak in [32,33].
1.5 Objectives
As Fuller says in [34], driving has, compared to most other things we do, a high
potential for adverse consequences. With an increasing amount of preventive safety
systems and dynamic adjustability of vehicles, knowledge of driver-specific differences
in preferences and performance can be of great use. If objective parameters correlated
to driving skill could be made available, this might be used for driver type identification
and this should enhance the benefit of adaptive configuration of adjustable systems.
Such parameters can also be used for metrics to model drivers of different levels of skill
for use in a desktop simulation environment, thus improving the validity of simulations
that depend on realistic driver models. Miller et al point out in [35] that intelligent
driver support systems with the human in control is a promising application area for
artificial intelligence research, with autonomous vehicles as a goal when such systems
have become a practical solution.
The presented research work is based on the assumption of a human remaining in
control most of the time, but with support from active, preventive and personalized
systems to improve performance. From the background described here, the objectives
of this research are summarized as: to define driving skill (and limitations to apply
here) as well as the recruitment process for finding test subjects from populations with
different driving skill; develop methods for characterization of the specified driving
skill; find objective measures for the specified driving skill; gather driver metrics, and;
develop a driver model that can represent different driving skill. The result is intended
to be a basis for future development of active safety system solutions that consider
different levels of path tracking driving skill.
1.6 Outline of thesis
This thesis comprises an introduction to the research topic, an overview of the
performed research, and five appended papers, Papers A – E. These papers include a
5
6
Chapter 1 Introduction
number of driving simulator tests designed and used for the study of driving behaviour.
Moreover, a test subject recruitment process has been designed to find drivers with
specific skills. The results from the simulator tests with these drivers are tested with a
method designed for identifying typical characteristics for different driving skill levels.
Driver model development has been carried out with specific focus on using a small
number of adjustable parameters that can be set for different path tracking skill levels.
The results from the driving simulator tests have been used for specification and
validation of the driver model.
Chapter 2 presents the method in more detail with the work process of the project,
including the limitations imposed. Chapter 3 gives an overview of the vehicle operator,
including descriptions of general human behaviour, the driving process and how
behaviour is shaped through experience. The purpose of this overview is to put the
work in the appended papers in a broader perspective and to describe the background to
and motivation for this research work. A brief introduction to measurements and
analysis of driver behaviour is given in Chapter 4. Chapter 5 describes the controlled
vehicle, both as a model interacting with the driver or the driver model, and as a model
that describes the driver’s interpretation of the controlled vehicle (i.e. vehicle model),
also used in the driver model presented in Paper D. Chapter 6 gives an overview of
different basic driver model types used for modelling individual drivers, as well as a
brief description of the developed model presented in Paper D. The main scientific
contributions are summarized in Chapter 7, the discussion and conclusions in Chapter
8, and the recommendations for future work in Chapter 9.
Chapter 2
Method
This chapter gives an overview of the work process of this research, with specification
of limitations and methods used, and provide information regarding the execution of
the experiments.
2.1 Work process
The first part of this work of finding objective parameter measures to characterize
driving skill, and to model drivers, was to find the right approach to the problem
through extensive multidisciplinary literature studies. These were conducted in a variety
of areas such as human cognition, psychology, driving behaviour, man-machine
interfaces, vehicle dynamics, driving tests, driving simulators, control theory,
optimization, pattern recognition and driver-vehicle modelling. Also, as part of the
background, field studies were done for vehicle handling skill and different vehicle
designs, including driving with different human vehicle interfaces (such as joystick
steering, variable steering wheel gear ratio and 4-wheel steering).
The first driving simulator experiment was scheduled quite early in the project with the
main objective of gaining experience, but also to test some initial hypotheses. The
experiment was executed in collaboration with a partner project, using the same
recruited test subjects for two different tests. The first test, done exclusively for the
other project, concerned passing of a truck with meeting traffic in other lane, while the
second part, done primarily for this project was designed as a curving road manoeuvre
at relatively low speed (using cones marking the way similar to some roadwork
situations). More information regarding the driving scenarios can be found in Section
4.3. Test subjects were recruited as being high mileage drivers (driving more than 25
000 km/year) or low mileage drivers (driving less than 5 000 km/year). Some of the
high mileage drivers were especially selected professional drivers, while the other test
subjects were recruited either from KTH or VTI. After testing and verifying the
simulator and scenario setup, a pilot test was performed using both low mileage and
high mileage drivers. This verified that the setup was right, allowing the full scale test
with 9 low mileage and 9 high mileage drivers being done (during 3 days, i.e. with 6
7
Chapter 2 Method
8
drivers each day). Each driver was tested in the two scenarios which took
approximately half an hour each to complete (the driver was allowed to re-run tests
which resulted in some variation in time). Both scenarios included an initial period of
test driving, and a break of half an hour was scheduled between the scenarios, in which
the drivers had to do some paperwork. The results from the first experiment resulted
both in Paper A and valuable input to the rest of the work in the project. Figure 3
presents an overview of the research work in this thesis.
Driving Simulator Test 1
LITTERATURE STUDIES / FIELD TESTS / EXPERT KNOWLEDGE
CONE TRACK (AVOIDANCE MANEUVRE)
• Driver tracking path differences
• Driver robustness for steering changes
• Driver repeatability
• Driver adaptation to steering changes
PAPER A
RECRUITMENT CRITERIA
Driving Simulator Test 2
PAPER B
DLC
ISO 3888-1:1999
Manoeuvring skill
LINE JUMP
Driver-vehicle
response
CURVING ROAD
Driving behaviour
on natural road
CURVING ROAD
Driving behaviour with
short sight distance
PAPER C
PAPER E
DRIVER METRICS
Driver Model Simulations
OPEN-LOOP
Vehicle data
for driver model
DLC
ISO 3888-1:1999
Manoeuvring skill
PAPER D
Figure 3: The work process with background studies, simulator studies, modelling, simulations
and analysis, and the resulting papers.
The second simulator test was designed exclusively for this work, studying different
approaches to find differences between low skill and high skill drivers. Definitions of
driving skill was formed based on expertise experience and state of the art descriptions
of driving, where path tracking skill was decided to be the main focus. These results
were used as basis for both the scenario descriptions and the recruitment criteria for the
test subjects. Drivers were recruited from populations described as low skill and high
skill drivers, with a selection of 15 drivers from each category. The selection of test
subjects was aiming to end up with a mix of both young and old, and men and women,
Section 2.2 Limitations
but with the main objective to find representative drivers for the two categories. The
preparation of the second simulator test was done in a similar way as the first, with
initial testing and a pilot test with representative drivers. Four scenarios were used,
divided into two sessions of approximately 25 and 19 minutes respectively. The first
two scenarios, a double lane change and a line tracking scenario, included a short
period of test driving. Between the two sessions was a break during which the test
subject answered questions about the tests. The first scenario served as verification of
the recruitment, testing the highest speed the driver could drive the manoeuvre, as well
as being a part of the regular measurement and analysis done on all scenarios. The
session after the break consisted of two scenarios based on driving on curving road.
This aimed to test the drivers in normal driving, with the third scenario using two levels
of road friction and fourth one using fog to test the effect of limited sight distance.
Three of the (low skill) drivers felt tendencies to motions sickness in the last scenario,
but these test subjects were removed from the analysis of that particular test. Thus the
analysis was based only on the remaining 27 drivers, instead of 30 for this scenario. See
Paper B, C and E for more details regarding the analysis of the scenarios in the second
diving simulator test.
The driver model was developed as the last part of the research work, based on the
results from the literature study together with the results from the driving simulator
tests and open loop simulations. The model was specified and tested with the double
lane change scenario measurement results from the driving simulator tests, with both a
high skill and a low skill setting specified based on the earlier results. See Paper D for
more details.
2.2 Limitations
In this work, methods have been derived to specify, identify, analyze and use
information regarding driving skill. The scope here has been limited to path tracking
driving skill of drivers holding a Swedish driving license, studying characteristic
behaviour of the group, i.e. not the individual driver. Only a few scenarios have been
used, which limit the direct applicability on identification of individual drivers in real
driving cases. However, the scenarios have been chosen in order to test behaviour that
has been considered most important for this work, representing an avoidance
manoeuvre, vehicle control, and driving on a curving road in some different conditions.
The study has also been limited to the driving simulator environment, and although the
experience is very similar to driving with a real vehicle, it does not completely replace
the need for such studies. Objective measures taken here are used for both
characterization and modelling typical behaviour of the two groups of drivers.
However, large experiments, at least partly done with real vehicles, should be
conducted before application in vehicle design or serial production of cars, or in-vehicle
identification of drivers, is feasible. Other types of skill could be possible identify, and
other scenarios could also be studied as well as the effects of other driving conditions.
Also, larger groups of test subjects could help in identifying other sub-groups of drivers
with statistical significance. It is also likely that the driving behaviour is different in
different parts of the world, which calls for studies using drivers from the target
markets.
9
10
Chapter 2 Method
Chapter 3
The human vehicle operator
This chapter gives an introduction to general human characteristics and more specific
driving-related processes, and provides a description of how experience can affect
performance and shape the behaviour of drivers. These individual differences in driver
characteristics are important for both the identification and the modelling of drivers.
3.1 General driver characteristics
This section includes an introduction to the purpose of actions, the basic properties of
humans, and the perception of information that is used when driving. Groeger stated in
[36] that actions are not simply responses or movements that have been initiated, but
also attempts to interact with objects in our world. Our movements are constrained by
the limitations of our limbs and the brain that controls them, but actions are also
constrained by the objects with which we interact. The process of a driver perceiving a
situation and responding with an action is very complex and involves many different
parts of the brain, from the recognition of the situation (e.g. through visual or tactile
stimuli) and taking a decision on which strategy to use for the task, to the very
coordination and movement of the limbs.
When trying to understand and model the human driver, a good starting point is to
study the general characteristics of the driver from a controller’s point of view. Kinecke
and Nielsen summarize the general characteristics of human controllers in [37] as
including:
•
Operating states (error correction mode, error ignoring mode)
•
Nonlinearity (e.g. accelerator pedal)
•
Adaptation (requires precise feedback)
•
Anticipation (learned behaviour, skill)
•
Time variance (time-dependent output)
11
Chapter 3 The human vehicle operator
12
A more thorough description of the physical limitations and attributes can be found in
[27], where MacAdam considers physical limitations such as human time delays and
threshold limitations. He also describes physical attributes such as preview utilization,
adaptive control behaviour, and the driver’s understanding of the dynamics of the
vehicle. Groeger has provided a source of information regarding driver characteristics
and aspects of human cognition in [36], and McRuer (e.g. [20]) is one of the most
important contributors to describing the human with a control-theory-based approach.
With regard to information acquisition, i.e. perception, MacAdam ranks the primary
sensory channels used in driving in the following order, with the first as the most
important: vision; the vestibular (inner ear) and kinesthetic (body-distributed) channels;
the tactile channel and; the auditory channel. Most information from cues with lower
priority also provides redundant/reinforcing cue information that can help the driver to
confirm decisions quicker and make better estimations based on information obtained
from the channels with higher priority [27].
Humans cannot process all the information available with the same level of detail. In
human information acquisition, the intake and processing abilities are limited and
queuing theory is one way to present a reproduction of the human system of handling
information. Queuing policy is a means of determining which information (clients) is to
be processed first and which has to wait or be dismissed. Normally the principle of first
come first served is used, but there is a range of other possibilities of handling
information [37]:
•
Priority Queuing
Clients with the highest priority are served first, independent of the arrival
order in the queue.
•
Limited Capacity Queuing
This involves a queue that can only hold a limited number of clients prior to
processing. No new clients are taken into the queue if it is full, and the clients
that have been in the queue too long will be removed without processing.
•
Pre-emptive / Non-pre-emptive
Clients appearing in the queue that has the higher priority than the one
currently being processed can result in two scenarios. The pre-emptive system
processes the higher priority client immediately and the client currently being
processed has to wait for service until the prioritized client has been served. A
non-pre-emptive system does not process the higher priority client until the
currently processed client has been served.
It is quite straightforward to apply these processes on automobile driving, which
requires adequate vehicle control combined with the performance of many other tasks
more or less loosely related to driving.
Section 3.2 The driving process
3.2 The driving process
Here descriptions of driving goals and how driving tasks are organised in the process of
driving are presented. For the purpose of determining what the driver actually is doing,
one approach is to study the choices of action which a driver has and what the driver’s
goals might be. Rumar has specified what he considers the driver’s high-level goals to
be in [38], maintaining that the primary goal for a driver is to reach his or her
destination. (Note that for some drivers the driving itself is the goal.) However, Rumar
points out that the driver will not accept that goal without reservations. The driver
requires a certain time expenditure, safety, economy, and comfort, and this is what
Rumar refers to as the driver’s secondary goals. He also stresses the necessity to
transform these general goals into operational goals. It is also important to note that the
driver behaviour cannot be isolated from the controlled system, since the driver can
compensate for relatively large differences in vehicle characteristics to achieve the
preferred system characteristics.
3.2.1 Driving functions and tasks
When studying operational goals, terms such as functions and tasks are introduced.
Functions can be seen as available functionalities in the joint driver-vehicle system that
enable the fulfilment of goals; and tasks can be regarded as actions that can be
performed with the help of those functions. In [36] Groeger exemplifies functions with
steering, speed control, gear changing, interpreting the road ahead, navigation etc.
Therefore, the driver task workload can be reduced by the automation of driving
functions. Since driving is a complex task, it requires a large set of activities to be
performed by the driver in a safe manner, but many types of activities can be described
as distractions from the main task (although sometimes these activities may increase the
driver’s general attention). Driver distraction is not in focus in the work conducted here,
and instead distractions are kept to a minimum to focus on the driving task. However, it
is necessary to have some insight into the topic and a great deal of information about
driver distraction can be found in [39]. A popular way of ordering driving tasks that has
been used with some differences in the choice of words, but with the same basic
categories, is to divide the tasks into a hierarchy of three performance levels, according
to the timescale and the level of cognitive activity involved (e.g. [40]):
•
Navigation (macro-performance)
•
Guidance (situational-performance)
•
Control (micro-performance)
Note that the levels in the model above concern tasks during the actual driving and not
tasks performed before or after driving, but strategic planning can be considered as
macro-performance as well.
13
Chapter 3 The human vehicle operator
14
Kiencke has made a task description of driving that is divided into more detailed levels
[37]:
•
Strategic tasks (choice of route, time of departure)
•
Navigational tasks (adherence to the chosen route during travel)
•
Traffic-related tasks (interacting with other road users in such a way that the
traffic is not obstructed and collisions with other road users are avoided)
•
Adherence to rules (traffic signs, signals etc.)
•
Tasks related to the road (chosen position within traffic, course)
•
Speed control (choice of speed according to the situation)
These levels are related to the previously described model, as the first two are
components of navigation, the third and fourth are components of guidance, and the
last two are components of control.
A complete driver model should, in addition to addressing activities on all of the levels
within the scope of the model, also describe the flow within the system. Figure 4 shows
the model by Repa, illustrating driver-vehicle interactions using the level description
above, with the addition of the communication, comfort, & entertainment level and the
pre-/post-driving level of accommodation [41].
Inputs:
• Driver goals
• Road conditions
• Traffic conditions
entering, positioning, securing, accessing, exiting
secondary display & control interaction
ACCOMMODATION
NAVIGATION
DRIVER
COMMUNICATION,
COMFORT &
ENTERTAINMENT
GUIDANCE
Outputs:
• Fit/position/comfort
• Navigation plan
• Path decision
• Path control
CONTROL
VEHICLE
lane position & speed
path selection
route finding & following
Figure 4:
Multi-level structure of driver-vehicle interactions [41].
Wheeler et al have made a very thorough description of driving that divides driving into
private vehicle operations and commercial vehicle operations [42]. The private vehicle
Section 3.2 The driving process
operations are listed below, with the pre-drive and drive tasks organized into separate
groups:
Pre-drive tasks
•
Inspection
•
Start-up
•
Auxiliary systems
•
Planning
Drive tasks
•
Navigation & routing
•
Guidance & manoeuvres
•
Control
•
Vehicle systems operation & monitoring
•
Reacting to emergencies
The pre-drive tasks are more or less covered by the accommodation level in the model
by Repa above, as well as by the initial strategic setup level of communication, comfort,
& entertainment and the navigation (i.e. planning) level. The drive tasks are here
extended with the vehicle systems operation & monitoring level (partly related to
communication, comfort & entertainment) and the reacting to emergencies level. This
allows a better overview of how the human-machine-interface (HMI) design affects the
driver, and provides the possibility of treating extraordinary events separately. In the
present work, the focus is on the behaviour of the driver when he or she is driving, i.e.
the drive tasks. Below the different sub-tasks of driving from [42] are described in more
detail, with the actual goals of the tasks within parentheses:
Navigation & routing
•
Way finding (going where one intends to go)
•
Route modification (changing the route based on the conditions)
Guidance & manoeuvres
•
Traffic coordination (keeping the vehicle at a safe distance from other
vehicles)
•
Rule compliance (keeping the vehicle within the regulated safety limits)
•
Manoeuvring (causing the vehicle to go where it is intended to go)
•
Hazard observation (avoiding hazardous situations)
15
Chapter 3 The human vehicle operator
16
Control
•
Speed control (matching the vehicle speed with the driving requirements)
•
Position control (matching the vehicle direction with the driving requirements)
Vehicle system operation & monitoring:
•
Monitoring engine operation (ensuring that the engine is operating normally)
•
Monitoring control systems and vehicle structure (ensuring that the tyres,
brakes, steering, and vehicle structure are functioning normally)
•
Adjusting climate control (maintaining a comfortable interior and clear
windscreen)
•
Initiating turn signals (warning other drivers of one’s turning intentions)
•
Operating communications systems (obtaining information from others and
giving information to others)
•
Using advanced traveller information systems (obtaining and using traveller
information)
•
Operating cruise control (reducing fatigue by automatic maintenance of speed)
•
Operating lighting systems (illuminating the highway and improving the
visibility of the vehicle to others)
•
Operating windshield washers/wipers (keeping the windshield clean)
•
Adjusting rear view mirror (providing a view of the traffic behind without
inducing glare)
Since reacting to emergencies is more of a sub-task itself (although not part of nonobstructed driving), no sub-tasks have been described for this level. However, as is the
case for all the other sub-tasks, there are even more detailed sub-levels described in
[42]. For reacting to emergencies, for example, these detailed sub-levels consist of the
following sub-sub-tasks:
•
Detecting emergency condition (detecting an emergency in time to take
corrective actions)
•
Diagnosing the situation (understanding what is happening)
•
Determining the action required (planning actions to mitigate the emergency)
•
Taking the appropriate action (mitigating the emergency)
For further understanding of why the tasks are grouped as they are, it can be mentioned
that many of the sub-tasks in vehicle system operation & monitoring are operations
performed on the same system functions as those affected by the sub-tasks of pre-drive:
auxiliary systems, i.e. functions not directly related to the level of control, but
performed in order to be able to operate the vehicle successfully in the desired way.
A slightly different approach to the components of driving is the functional model
called Driver in Control (DiC) presented by Hollnagel et al in [9]. It describes driving
Section 3.2 The driving process
as cycles that link intentions/objectives, actions, and outcomes together between
different levels. The vehicle systems are here considered in the same framework as the
driver, and treated as an integrated system. Every level contains a cycle, but with
different characteristics in terms of the type of control (feedback or feed-forward), for
example. The different loops in the model are referred to as the tracking, regulating,
monitoring, and targeting loops. Figure 5 shows the communication downstream
between the loops, and also what is related to closed loop compensatory control and
open loop anticipatory control.
Figure 5: Principles of the DiC model (with the feedback links and unexpected events removed
from the figure to avoid cluttering) [9].
The different loops are presented below in more detail, beginning with the lowest level
(tracking loop) since these activities are the basic ones required for short-term
controllability.
•
Tracking loop
The tracking level describes the low-level activities required to keep the
vehicle inside a region of time-space continuum, i.e. to maintain the speed, the
distance from the car in front (and behind), the relative or absolute lateral
position etc. These activities are described as mainly closed loop control,
which skilled drivers can accomplish with little effort and without paying
much attention to them. If the conditions change, however, the actions may
require more attention and thereby become more like those at the regulating
level. The typical duration of activities in this loop is less than one second.
•
Regulating loop
The regulating level provides the input (goals and criteria) for the tracking
level. The activities at this level concern mostly closed loop control, although
some anticipatory control may occur. They include activities such as
regulating the target speed, the specific position and the movement relative to
17
Chapter 3 The human vehicle operator
18
other traffic elements, as well as the state of the joint driver-vehicle system
relative to the driving environment (traffic flow, hazards). These activities
refer to specific plans and objectives coming from the monitoring level. One
activity at this level is to generate plans and objectives for the tracking level,
and a number of tracking sub-loops may be involved. This requires a higher
level of attention than the tracking level. The typical duration of activities in
this loop is between one second and one minute.
•
Monitoring loop
This level concerns the monitoring of the joint driver-vehicle system, e.g. the
vehicle status, location, available resources, etc. Monitoring also concerns
keeping track of traffic signs and signals such as indications of directions,
warnings (e.g. road conditions or curves), and restrictions (e.g. speed limits).
Monitoring is therefore a mixture of open loop and closed loop control.
Modern vehicles have systems that only warn in cases of severe malfunction
and systems that help to clarify ambiguities. The typical duration of activities
in this loop is from 10 minutes to the duration of the journey.
•
Targeting loop
Targeting is distinctly an open loop activity, implemented by a non-trivial set
of actions, often covering an extended period of time. The activities at this
level are mainly carried out prior to the journey, but events during driving may
trigger an activity, e.g. changing the route, driving faster, etc, i.e. targeting
activities. The trigger can be a landmark, traffic information etc. Assessing the
change relative to the goal is not based on simple feedback, but rather a loose
assessment of the situation, e.g. the estimated distance to the goal. When these
activities are performed regularly, they may be considered a part of
monitoring. The typical duration of activities in this loop is up to a few
minutes.
It is essential that the joint driver-vehicle system (JDVS) should be in control at all the
levels all the time for the system to have an effective control. Ineffective control means
that control has been lost in one or more of the loops, thereby risking unwanted effects
on the joint driver-vehicle system. One advantage mentioned for the DiC model is that
it explicitly describes how disturbances can propagate between control levels to several
levels, and not just how an activity at one level can affect activities at another level. For
example, if the required time of arrival at the destination is changed in the targeting
loop, this will propagate to the monitoring loop, where the monitoring priority may
change. Moreover, the new plans may result in more risky manoeuvres being chosen in
the regulating loop, and this may result in active safety systems kicking in and thus
affecting the driver-vehicle control in the tracking loop. The driver and vehicle are in
most cases treated as a combined system, referred to as the driver-vehicle system [14],
which in many instances cannot be separated into mechanical and purely human
components due to the highly coupled interaction [27].
Section 3.3 Individual experience-based differences
3.2.2 Specification of driving activities
Based on the descriptions above, it is defined here how the activities are grouped for the
present research work. They are divided into three main categories depending on their
relation to the driving referred to as path tracking, i.e. following a chosen driving path:
•
Primary driving functions/tasks
Activities at the level of guidance and control, i.e. the activities required by the
driver (and/or driver support systems) to keep the vehicle on the road and keep
the distance to other objects.
•
Complementary driving functions/tasks
Activities related to navigation as well as activities performed to
maintain/improve driving conditions (keeping the windshield clean, etc).
•
Non-driving functions/tasks
Activities not intended to improve driving (talking on the phone, yelling at the
children, adjusting the radio, etc).
Since the research work presented here is focused on path tracking activities, this limits
the scope to the first category of the three. This does not mean that the other functions
do not influence the driving, but the aim of the experiment setup is to reduce the need
for such activities to a minimum. It is also the undisturbed actions of the driver that are
the main concern, and therefore the primary driving tasks of the driver are kept free
from interference from preventive safety systems within the primary driving functions.
3.3 Individual experience-based differences
Here theories are presented on how experience shapes human behaviour and how it may
improve the performance in certain situations. There are hypotheses concerning how
human cognition works which are difficult to combine, but which are allowed to coexist since the different cognitive processes are difficult to verify or reject. Theories
live and prosper in parallel, with the domination of some of them. Therefore, it is
difficult to obtain a complete overview of the state-of-art within the area, but here is an
attempt to provide an objective introduction to the topic.
3.3.1 Behaviour development
There appears to be general agreement concerning the postulate that most human
behaviour is goal-oriented, at least when we make conscious decisions. Rasmussen [43]
is of the opinion that human activity in a familiar environment will not be goaloriented, but rather controlled by a set of rules that has been proven successful
previously. When performing a task like driving, we have several sub-tasks that we
have to do which are more or less an act of routine, and require less conscious thought
19
Chapter 3 The human vehicle operator
20
than other sub-tasks. Without an explicit or implicit plan, however, no goal is achieved.
Actions are controlled by intentions, but all intentions are not carried out. Some are
abandoned and some are revised to fit changing circumstances [44]. It is also important
to realize that it is not the physical reality that decides behaviour, but the perceived
information. Every road user selects his or her own information [38].
Many scientists base their work on Rasmussen’s three-level behaviour model with its
mapping of driving tasks to human cognitive levels [43]. The three levels are:
•
Skill-based behaviour
•
Rule-based behaviour
•
Knowledge-based behaviour
An illustration of the model can be seen in Figure 6.
GOALS
KNOWLEDGE- SYMBOLS
BASED
BEHAVIOR
RULEBASED
BEHAVIOR
SIGNS
IDENTIFICATION
DECISION
OF TASK
PLANNING
RECOGNITION
ASSOCIATION
STATE/TASK
STORED RULES
FOR TASKS
SKILLBASED
BEHAVIOR
(SIGNS)
FEATURE
FORMATION
AUTOMATED
SENSOR-MOTOR
PATTERNS
Figure 6: Simplified illustration of the three levels of performance of skilled human operators.
Note that the levels are not alternatives, but interact in a way only rudimentarily represented in
the diagram [43].
•
Skill-based behaviour
By skill-based behaviour, Rasmussen refers to sensory-motor performance
during acts or activities which, following a statement of an intention, take
place without conscious control. He describes them as smooth, automated, and
highly integrated patterns of behaviour. Performance is rarely based on simple
feedback control where motor output is a response to the observation of an
error signal representing the difference between the actual state and the
intended state in a time-space environment. Skilled behaviour comes from the
ability to compose sets of automated behaviour from previous experience for
use in a specific situation.
21
Section 3.3 Individual experience-based differences
•
Rule-based behaviour
At the next level, rule-based behaviour, the composition of a sequence of
subroutines in a familiar work situation is typically controlled by a stored rule
or procedure which may have been derived empirically on previous occasions,
communicated from other persons’ know-how as an instruction, or prepared on
the current occasion through conscious problem solving and planning.
•
Knowledge-based behaviour
Knowledge-based behaviour is what we are forced to use in unfamiliar
situations, faced with an environment for which no know-how or rules for
control are available from previous encounters. At this level, performance is
goal-oriented and based on knowledge.
According to Goodrich et al, in driving, human cognition can be described using
multiple mental models (agents) which can be organized into a society of interacting
agents whose structure determines both which agents contribute to driver behaviour and
which agents can employ attention resources [45]. As can be seen in Figure 7,
Rasmussen’s behaviour model [43] is used as a starting point for further exploration of
the concept of mental models. The three levels are skill-based (SB), rule-based (RB),
and knowledge-based (KB) behaviour. The figure uses the notation SP for sensor
perception, MM for mental model, and BA for behaviour actuation.
KB
SP
MM
BA
RB
SP
MM
BA
SB
SP
MM
BA
WORLD
Figure 7:
Communication and control within a society of mental model agents [45].
An important distinction in Goodrich research work is a coordination that describes
when a driver switches between different skill-based (SB) agents and how attention is
shared between agents (Figure 8). Each mental model is assumed to be either enabled or
disabled, and engaged or disengaged. With an enabled mental model there is active
influence on human behaviour generation and with a disabled mental model there is no
such active perception occurring. When a mental model is engaged, attention is being
maintained, whereby environmental information is actively being perceived and
interpreted, and when the model is disengaged, attention is released and no such active
perception occurs.
Chapter 3 The human vehicle operator
22
SUPERVISORY
AGENT
KB BEHAVIOR
LONGITUDINAL
VEHICLE
CONTROL
RB BEHAVIOR
SB BEHAVIOR
sr
tr
LATERAL
VEHICLE
CONTROL
USE
CARPHONE
ba
Figure 8: Hierarchical structure of agents in a mental society. SB behaviour is exemplified by
speed regulation (sr), time headway regulation (tr), and braking to avoid collision (ba) [45].
It is assumed in the theory by Goodrich et al that attention cannot be divided, but must
be switched between rule-based (RB) behaviours in such a way that knowledge-based
(KB) agents plan and coordinate RB agents. The RB agents determine which SB agent
to enable, when to switch from one SB agent to another, and which sensors should be
consulted to reduce uncertainty and ensure satisfactory performance. To disable one
agent and enable another, the RB agent must identify when currently enabled SB agents
cannot accomplish the assigned RB task.
SB controllers execute the task specified by the RB agent, e.g. speed regulation (sr),
time headway regulation (tr), and braking to avoid collision (ba). High workload and
high perceptual bandwidth tasks must be allocated high attention, and therefore such
criteria are communicated within the structure. In these situations the previously
described queuing policy by Kinecke and Nielsen [37] plays an important part.
The task demand is defined by Fuller as the objective complexity of the task and arises
out of a combination of features of the environment, the behaviour of other road users,
the control and performance characteristics of the vehicle, its speed, road position and
trajectory, and driver communication, see Figure 9 [46].
TASK DEMAND (D)
ENVIRONMENT
ROAD POSITION
AND TRAJECTORY
COMMUNICATION
SPEED
Figure 9:
Task demand, by Fuller [46].
OTHER ROAD USERS
VEHICLE
Section 3.3 Individual experience-based differences
3.3.2 Driving skill and driver performance
How should one relate the general term driving skill to the driver and his or her
behaviour? It may be intuitive for some to relate this skill to the actual control of the
vehicle, i.e. the precise positioning of the vehicle on the road. Others may primarily
relate issues regarding how accurate a person is in predicting and avoiding hazardous
situations to driving skill, and some may consider driving skill primarily to be a
measure of how accurate the driver is in driving by the traffic regulations. One could
say that a skilled driver would achieve a specific goal in a better way than an unskilled
driver in all these examples, but the question arises as to whether the definition of the
goal is detailed enough for both the observer and the driver to have the same
preferences when interpreting what ‘better’ means. Jarlmark deals with the problem of
specifying which control behaviour is preferable in a situation, which he asserts is
exemplified by the fact that a quicker and more accurate response might actually put the
vehicle closer to the limit of the vehicle manoeuvring [47]. In [37] Kiencke refers to a
definition of a good, safe driver as follows:
•
Has complete command over the vehicle equipment
•
Keeps fairly well to the rules
•
Takes no unnecessary risks
•
Has good ability to anticipate the traffic situation in the immediate future
•
Shows consideration for other drivers when they make mistakes
•
Keeps his temper under control
Another aspect of skill which is shown by Mitschke [48] and which involves a
controller’s viewpoint rather is the ability of the driver to adapt or tune himself to the
driven vehicle, which mathematically expressed means that the parameter values in the
driver equation are dependent on the parameters in the vehicle equations. This again
shows the need for treating the driver and the vehicle as an integrated system. In the
research work presented in Paper A, the drivers were tested for their ability to adapt to
changed steering characteristics, which some drivers managed to do almost
instantaneously. Elander et al [49] refer to skill as the ability of drivers to maintain
control over the vehicle and respond adaptively to complex driving situations, or what
they express in other words as driver performance; and they state that driving skill is
expected to improve with practice or training.
Regarding potential skill, Groeger in [36] suggests that performance, as the driver
gradually gains more practice, does not reflect the improvement of single components,
but rather the grouping (or “chunking”) of performance components, and consequently
skilled behaviour is based on a larger organization of components than unskilled
behaviour. He also argues that people develop a measure of proficiency at a number of
skills, rather than one single skill, “driving”, and continues by stating that there also
exists a transfer of what is learned in one situation to another situation, to a certain
degree. It is possible, according to Groeger, that such transfer occurs through
declarative similarity rather than functional similarity; i.e. the transfer of skill from one
driving task to another is quite limited between different tasks, e.g. turning right
23
Chapter 3 The human vehicle operator
24
compared to driving straight ahead in a specific junction, while the rate of learning a
specific task in different situations, e.g. turning right at different types of junctions, is
quite high. Thus driving requires not only multiple functions (e.g. steering, speed
control, gear changing, interpreting the road ahead, navigation etc.) to be taken into
account, but also the performance of a continuous stream of tasks that are quite
different from one situation to another.
Another distinction to be elucidated is the difference between potential and effective
skills in driving. Fuller uses the term capability to refer to the momentary ability of the
driver to deliver his or her level of competence, i.e. what the driver actually is able to
do at any moment, as seen in Figure 10 [46]. By competence he refers to the driver’s
attainment in the range of skills, referred to as roadcraft in his paper. Roadcraft is
described as a concept which includes control skills, the ability to “read the road”
(hazard detection and recognition), and anticipatory and defensive driving skills. This
way of viewing capability and competence can be of use when considering the issue of
potential and actual performance, which are terms that are very relevant to this study,
especially when designing the driving experiments.
CONSTITUTIONAL
FEATURES
TRAINING & EDUCATION
EXPERIENCE
COMPETENCE
HUMAN FACTORS
CAPABILITY (C)
Figure 10: Determinants of capability, by Fuller [46].
There are some different opinions about what actually happens when humans perform a
familiar task/observation, i.e. when we practise what one could call an acquired skill.
Some researchers support a theory of automatic behaviour in combination with the
concept of a central bottle neck, maintaining that a certain level of practice can result in
an automated behaviour that will release attention resources for other tasks [50]. Others,
like Neisser [51], argue in favour of the theory that we create schemata for different
tasks, with practice refining the schemata to cope with new situations, which at first
require more attention, during training, but then less attention when the schemata are
learned well. Neisser argues that, even if we are unaware of ourselves performing
actions, this does not mean that we are automatically controlled by simple stimuli.
Neisser questions whether any responses but primitive pre-cognitive actions can ever be
really automated in the sense that they can be performed no matter what the situation is
and without the influence of the person’s own plans and intentions. Groeger also argues
that there are reasons to think that activities like braking, car following, and curve
Section 3.3 Individual experience-based differences
driving require more selective and deliberate action than is possible without attention
[36], i.e. in the ‘automated’ way. However, the difference between the two theories is
not critical for the work in this thesis. The main observation is that, with either of the
two theories (automatic behaviour or schemata), more resources can be allocated to
secondary tasks if the driver is familiar with the situation of the primary task.
Wilde presents in [52] a behaviour theory covering issues regarding the risk-taking
behaviour of drivers, the theory of risk homeostasis. The theory is well known and the
object of debate, and can be summarized as stating that we prefer to adjust our
behaviour so that the perceived risk will match the level of target risk, and thereby
eliminate the benefit of introduced risk countermeasures. Michon is somewhat critical
of the theory of risk homeostasis and maintains: “Only on the extreme implausibility
and much too strong assumption that the same homeostat is operating in all individuals
(rather than weakly, but more plausibly assuming that any human behaviour is adaptive
in some generic sense) is Wilde’s model theoretically correct” [53]. Regardless of the
correctness of the theory, Wilde in [52] also includes a useful definition of three types
of skills that exert an effect upon driver behaviour:
•
Perceptual skills (the extent to which the subjective risk corresponds to the
objective risk)
•
Decisional skills (the ability of the driver to decide what should be done to
obtain the level of risk that the driver prefers)
•
Vehicle handling skills (how effectively the driver can carry out his decisions)
All the three types of skills presented above may be improved by enhancing driver
education, licensing standards, and the ergonomic environment (e.g. the road geometry,
signals, and the controls and displays in vehicle design), according to Wilde. With this
definition of skill, all the activities, or performance levels, related to the actual driving
are included and therefore not only the low-level control; i.e. vehicle handling skills are
considered to be only one of three types of skill. All these levels of skill are treated in
the driver model in Paper D, although the perception of both high and low skill drivers
is assumed to be very good in the specific test used. The first type of skills, perceptual
skills, is in this thesis expanded with anticipatory and interpretational skills to
emphasise the driver experience-based ability to predict future states. This is similar to
the definition used in [5] for situational awareness, describing it as consisting of three
parts. The first part concerns perception of elements in the environment (e.g. the road,
traffic and vehicle condition), the second is understanding the current situation (e.g. that
a red light on the car in front is an indication of braking), and the third one concerns
foreseeing the near future, achieved through understanding the state and dynamics of
the elements around us (e.g. understanding that the distance to the vehicle in front will
be reduced if that vehicle is braking).
3.3.3 Driving style
It is not obvious how to draw a clear line between skill and style for a specific type of
behaviour, but with the previously described importance of a well-defined goal for skill
in mind, it can be concluded that, at a high level, style should be defined without any
25
Chapter 3 The human vehicle operator
26
relation to a specific goal. In Elander et al [49], style is referred to as the way in which
drivers choose to drive or habitually drive, including the choice of driving speed and
headway, and the habitual level of general attentiveness and assertiveness. Driving style
is expected to be influenced by attitudes and beliefs regarding driving, as well as more
general needs and values. In the previous chapter, Neisser’s theory of schemata [51]
was mentioned. If that theory is used, it could be suggested that a person’s own style of
driving influences every aspect of driving, even the lowest level of tracking/control.
If we study different classifications of drivers, Taubman-Ben-Ari et al subdivide the
driving styles found in literature into four broad domains [54]:
•
Reckless and careless (deliberate violations of safe driving norms, and the
seeking of sensations and thrills in driving)
•
Anxious (driver stress-related, and reflecting feelings of alertness and tension,
as well as ineffective engagement in relaxing activities during driving)
•
Angry and hostile (expressions of irritation, rage, and hostile attitudes and acts
while driving, reflecting a tendency to act aggressively on the road)
•
Patient and careful (planning ahead, attention, patience, politeness, and
calmness while driving, and keeping the traffic rules)
Continuing with Taubman-Ben-Ari et al, factor analysis is used for the specification of
a more detailed description with eight main factors representing specific driving styles:
•
Dissociative (e.g. making misjudgements, performing wrong actions,
forgetting)
•
Anxious (nervousness about driving, frustration, worrying, driving slow)
•
Risky (e.g. taking risks, driving at the limit, performing non-driving tasks such
as fixing one’s hair/makeup)
•
Angry (e.g. shouting, using the horn, flashing lights, manoeuvring to obstruct
others)
•
High-velocity (e.g. being impatient, tailgating, driving against a red light)
•
Distress reduction
meditation)
•
Patient (e.g. letting others go first, planning in advance)
•
Careful (e.g. driving cautiously, readiness for unexpected manoeuvres by
others)
(performing relaxation activities / muscle relaxation,
The driving style factors are in the paper coupled to driving styles and personality traits
like self-esteem, need for control, sensation seeking, and extroversion.
Many studies have been conducted focusing on driver behaviour when using adaptive
cruise control (ACC), and the classification of drivers in these studies can be in terms of
willingness to pass, distance to the vehicle in front, or similar traits (e.g. [55]).
Section 3.3 Individual experience-based differences
A commonly used classification of driving styles, exemplified by Tricot et al, is based
on the following general classes [56]:
•
Economical style
•
Normal style
•
Sporty style
These are distinguished by the variables accelerator and brake pedal position, steering
wheel angle, engine speed, vehicle speed and acceleration in the paper by Tricot et al. A
similar categorization was used for the subjective description of a system by
Barthenheier et al in [57], where it was applied for the evaluation of parameters for a
steer-by-wire system:
•
Comfortable
•
Sporty / encouraging active driving
•
Safe (subjective measure)
•
Generally appreciated
A system with specific characteristics can affect the driver behaviour, since drivervehicle interaction works both ways; for example a sporty feeling is connected to the
encouragement of active driving.
Tricot et al also refer to studies where driving styles are differentiated according to the
time interval used: a short or a long time interval. The influence of the environment has
been taken into account by a few studies, e.g. the road types (urban roads, A-roads / Broads, and motorways) and the traffic density (dense and light traffic) [56]. In addition
to the driving style classification above, the economical, normal and sporty styles,
Tricot et al choose to include the environmental variables dense and light traffic in their
study. Instructions were given to the drivers, instructing them as to which of the three
styles to use; e.g. for the sporty driving condition, drivers were instructed to “drive
rather fast but keeping good safety margins”, an instruction that could be interpreted
differently by different drivers. The study succeeded in making a fairly good distinction
between the styles. However, a couple of drivers were clearly classified into the wrong
group of style. This was considered to be partly a consequence of the individual
differences in their judgment of what was sporty, normal or economical driving. It is
important to describe such subjective terms such in more detail, preferably with
examples, since they otherwise are very likely to be interpreted differently for different
groups of people. Unfortunately the study also used only 13 subjects and a fixed based
driving simulator (with a real car cabin).
In the thesis by Jarlmark [47], a categorization of drivers was made according to their
compensation velocity, compensation precision, and driving strategy when
compensating for a crosswind gust. The drivers were divided into the following
categories: quickly and slowly compensating (QC and SC), over-, under- and mixedcompensating (OC, UC, and MC), and continuously oscillating and “correct and hold”
(CO and CH) drivers. All of these characteristics seemed to be independent and thus
resulted in twelve identified combinations that could be used for further study.
27
Chapter 3 The human vehicle operator
28
As a last remark about style, we can state once again what was mentioned earlier about
human behaviour: intentions can only be expected to predict a person’s attempt to
perform, not necessarily his or her actual performance, since external or psychological
factors may prevent the actions from taking place [36].
3.3.4 Specification of ordinary driving and path tracking skill
This thesis work concerns ordinary driving, which is defined here as driving whose aim
is to achieve safety, compromising performance and comfort. Given this condition, path
tracking driving skills are defined here as comprising:
a.
The ability to drive with lower friction and at higher speed (handle the vehicle
at the limit)
b.
The ability to recognize where the limit is
c.
The ability to drive with the lowest lateral accelerations and/or sideslip angle
(path-dependent)
d.
The ability to drive with low deviation from a driver-preferred track (driver
strategy)
e.
The ability to adapt to new vehicles/situations (always keeping a relevant
internal model)
f.
The ability to correct for unexpected disturbances
These different aspects of driving skill are in most respects covered in the different
papers presented here. Both a) and b) relate to at-the-limit handling, which is not the
specific focus of the present research work, but the double lane change manoeuvre in
Paper B includes driving that approaches the limit for higher speed. This manoeuvre
also benefits from skill related to c) and d). Paper A includes analysis of c), d) and e)
for drivers in general. Unexpected disturbances, f), are covered in Paper C, which
investigates the driver reaction to a sudden preview path movement. The effect of
external forces, such as side wind or tyre failures, may require additional driving
competence, and is not covered here (but studies are made in other projects in the
research group, see e.g. [47,93]). The driver model in Paper D concerns tracking a
path, which relates mostly to d), although it also covers the same aspect as that dealt
with in Paper B. The curving road driving in Paper E relates to c) and d) in general,
but for very short sight distances the sudden appearance of left and right curves
introduces an unexpected disturbance that requires some skill according to f).
Chapter 4
Measurement of driver
characteristics
This chapter presents methods for measurement and analysis of driver characteristics,
followed by a description of the driving simulator used for the testing and short
descriptions of the different simulator tests performed.
4.1 Introduction to driver analysis
There are two main purposes for measuring driver characteristics: to identify the driver
or driver type, and to model the driver or driver type. There are significant challenges in
modelling driver behaviour since much remains to be learnt about the structure, order,
or granularity of an individual’s control system. Human control strategy is both
dynamic and stochastic in nature, and the complex mapping between sensory inputs and
control action can be highly nonlinear [59].
Diagnosis methods for systems are commonly divided into two families [56]:
•
Internal diagnosis methods
Comparison between system model outputs and actual system outputs. A
sufficiently complete and precise model of the driver is required in order to
describe causality relationships between the information collected by the
driver and the reason for his actions.
•
External diagnosis methods
Methods based on observation of the inputs and outputs of the process, using
statistical analysis methods. These methods do not need any explicit model of
the actual system.
For the research presented in [56], Tricot et al could not find any driver model which
could deal with internal processes, which could run on a computer, and which was
29
Chapter 4 Measurement of driver characteristics
30
sufficiently accurate to be used in a model-based diagnosis application. Based on the
assumption that such a model did not exist, the second approach, an external diagnosis
method with factor analysis and pattern recognition, was used in their paper. Nechyba
et al [59] also argue in favour of the benefits of using an external diagnosis method,
since no explicit physical model is required, but they assert at the same time that the
lack of scientific justification of such learned models detracts from the confidence that
we can show in them. For a dynamic process, model errors can feed back to themselves
to produce trajectories which are not characteristic of the source process and may even
be unstable. For a stochastic process, a static error criterion based on the difference
between the training data and the predicted model outputs may be inadequate and
inappropriate for gauging the fidelity of a learned model to the source process. Nechyba
et al state that most learning approaches utilize some static error measure as a test of
convergence for the learning algorithm, but offer few, if any, guarantees concerning the
dynamic behaviour of the resulting learned model. Statistical error measurement
methods, such as the root-mean-square (RMS) method, do not provide sufficiently
satisfactory model validation for a dynamic process according to the authors, and
therefore the Hidden Markow Model (HMM) was used instead in [59] as a validation
method for the trained models.
For driver modelling with different levels of detail however, it should be noted that
external diagnosis methods used carefully can provide a powerful tool for analysing at
least higher-level tactical driver behaviour, since the complex decision-making of
humans does not have to be explicitly described, while detailed descriptions using
internal diagnosis can give the needed insight into low-level operational driving
behaviour.
4.2 Examples of measurements
According to MacAdam [55], it is good to have a thorough understanding of the range
of driving styles, including factors such as:
•
Acceleration / deceleration comfort levels
•
The headway gap sizes employed during following
•
The overall level of aggressiveness related to passing and overtaking activities
Savkoor et al perform driver strategy classification according to acceleration in [60],
which is illustrated in Figure 11.
31
Section 4.2 Examples of measurements
a)
b)
ax [m/s2]
CB
d)
c)
ax [m/s2]
Rmin
s [m]
e)
ax [m/s2]
CB
Rmin
s [m]
ax [m/s2]
CB
Rmin
s [m]
f)
ax [m/s2]
CB
Rmin
s [m]
CB
Rmin
CB
Rmin
s [m]
ax [m/s2]
s [m]
Figure 11: Examples of some elementary braking strategies while approaching and negotiating a
curved road segment, specifically in relation to the curve beginning (CB). a),b),c) Braking on a
straight road segment to the desired speed before the curve begins (pure longitudinal slip
followed by pure lateral slip). d),e),f) Braking on a straight road segment and in the curve to the
desired speed (pure longitudinal slip followed by combined longitudinal and lateral slip) [60].
In [61] Kuge et al describe a method for driver behaviour recognition that is based
“entirely” on HMM, with continuous recognition of driver behaviour. Multiple
correspondence analysis (M.C.A.) was used by Tricot et al in [56] to identify the best
set of variables to characterize the studied drivers’ behaviour. Discriminant analysis
(D.A.) was used afterwards for automatic classification of new observations. In [32,
33], Raksincharoensak et al evaluate the prediction of pattern recognition methods in
longer periods of normal driving.
How we perceive information through vision when we are driving is an important
parameter when classifying different drivers, since vision is by far the most important
sensory queue in driving, although the level of importance depends on the scenario and
the situation [27]. Inter-event arrival and service times are described in [37] by Kiencke
and Nielsen. The overall viewing frequency is divided between various viewpoints on
the road:
•
Visual focus (the point which is approximately a 3 second drive away)
•
Lead point (furthest point of the driver’s view)
•
Road edge
•
Road centre (i.e. central reservation)
Service time is the amount of time which the driver spends looking at a particular point,
which is given as between 0.25 s and 1.8 s for road viewpoints [37]. MacAdam lists
vehicle response signals presumed to be sensed by the driver model for steering and
speed control purposes [62]:
Chapter 4 Measurement of driver characteristics
32
•
Lateral acceleration
•
Lateral vehicle position
•
Longitudinal vehicle position
•
Vehicle heading angle
•
Vehicle forward speed
•
Vehicle lateral speed (sideslip velocity)
•
Vehicle yaw rate
•
Vehicle roll angle
•
Vehicle roll rate
These parameters are therefore to varying degrees a part of the driver’s perception of
the state of the vehicle in the model by MacAdam. From simulator runs described in the
same paper, and further analyzed in [27], one of the conclusions is that novice drivers
were more likely to drive slowly and sacrifice path accuracy to retain directional
stability near the handling limit, in comparison with expert drivers. The latter were
more successful in performing the required manoeuvres at higher speeds, but they were
also more likely to exhibit directional instability.
Underwood et al studied eye movement patterns in [63] and noticed a distinction
between drivers with various experience. They also referred to studies that showed
novice drivers to be focusing longer on hazardous objects and to detect fewer peripheral
events than experienced drivers. Piechulla et al [6] used a secondary task consisting of
reading a scrolling text in a pilot study to access the driver’s workload, which is
possible since the visual workload is considered a crucial component of the total
workload. The driver was instructed to consider safe driving as the primary task, and
that the driving performance should not suffer from the secondary task. In the study, the
number of glances per second was measured, since the glance frequency was described
as a sensitive measure of the driver’s visual workload.
Weir and Allen used an electrocardiogram (ECG) and derived the heart rate for
measuring the driver stress level in a variety of driving tasks, and concluded that this
method provides a comprehensive measure. The stress level is considered to be an
important measure of the driver’s state of alertness and level of skill in a given
situation. The heart rate is inversely related to the task stress, due to sinus arrhythmia
(the influence of breathing on the heart rate) [64].
In the research work here, the focus has been on parameters that are measured in the
standard vehicle hardware, i.e. not on parameters obtained by measurements performed
directly on the driver. In Paper A, the data was analysed for differences in driver
behaviour using only the driven path and the lateral acceleration levels in a repeated
guided manoeuvre. The path was analysed in three ways: using the average over a large
number of paths to study the difference between the drivers that could be related to
cornering strategy; comparing the individual drivers’ deviation from the mean path, i.e.
ascertaining the driver repeatability; and investigating the change in deviation from the
mean path using modified steering characteristics, i.e. the driver robustness for system
33
Section 4.3 Driving simulator tests
changes. In the following research work, Paper B to Paper E, a large number of
measures were taken from available vehicle parameters, including parameters with
reference to the road curvature as well as the steering wheel movements. Table 1 shows
the full set of parameters used in this work. This includes most of the response signals
listed by MacAdam in [62], with the addition of angular and torque information for the
steering wheel.
Table 1: Evaluated parameters.
Parameter
yroad
ψroad
ay
ψ
β
δSW
MSW
Description
Lateral position, road reference
Yaw angle, road reference
Lateral acceleration
Yaw angle
Body slip angle
Steering wheel angle
Steering wheel torque
1:st derivative
dyroad/dt
dψroad/dt
day/dt
dψ/dt
dβ/dt
dδSW /dt
dMSW /dt
2:nd derivative
d2yroad/dt2
d2ψroad/dt2
d2ay/dt2
d2ψ/dt2
d2β/dt2
d2δSW /dt2
d2MSW /dt2
The parameters has been analysed with relation to the test subjects’ driving skill, using
a number of measures taken in selected parts of the scenarios. More information on this
procedure is found in Paper B, where the method for evaluating the measures is also
described. This method is also used for the evaluation of measures in the different
manoeuvres and driving conditions described in Paper C and Paper E.
4.3 Driving simulator tests
For analysis of the driver behaviour, it does not suffice to analyse the driver isolated,
since the driver and vehicle form a coupled system where only the driver behaviour is
affected by the driven vehicle. For example, the driver will adapt to changes in the
steering wheel gear ratio (Paper A). Since experiments using real vehicles are very
sensitive to changes in the driving conditions and may be dangerous for the driver, it is
often preferred to use driving simulators instead. Besides that, simulator tests are more
time-efficient and can in many cases be more cost-efficient as well, especially if
different or complex vehicle and scenario setups are used. Human limitations in
remembering and comparing experiences can also be addressed with the instantaneous
system or scenario changes that are possible in the simulator.
Simulators have been used for both open loop and closed loop simulations. The closed
loop simulations need a strategy for the feedback control, which can be provided either
by a driver model or a living operator, i.e. a human driver. The human operator requires
an accurate feedback of the system behaviour to achieve realistic control. Fixed base
simulators are the most common type for human feedback control, and can be made
quite advanced, even with regular personal computer systems and other off-the-shelf
products. These simulators can provide visual, auditory and tactile feedback for the
driver, but are not equipped to excite forces to represent the movement of the vehicle.
This limits the possibility of achieving a realistic response from the driver, especially in
situations that normally would generate high vehicle accelerations. For that purpose
moving base simulators can offer a better platform for tests.
34
Chapter 4 Measurement of driver characteristics
Moving base simulators provide a realistic complement to real vehicle tests. The
repeatability and possibilities of dynamic changes can provide high quality data with
little noise. However, the quality depends on the realism of the driving experience and
how well different setups are represented by the simulator. A simulator test is only a
representation of a real test and the experience can not yet be made into an identical
substitute in all respects. Therefore, it is important that thorough real vehicle tests
should be performed before the introduction of innovations in production vehicles.
However, the results from driving simulator tests can significantly reduce the number of
tests to only the most promising configurations.
4.3.1 Test platform
The driving simulator experiments in the present research work were performed using
VTI Simulator III [65,66,67] (Figure 12), which has been built around a real vehicle cab
and utilises a sophisticated motion system which enables fast acceleration. This
simulator can be connected to a vehicle model in an MBS program such as CarSim,
which was used in this work. The CarSim model is further described in Section 4.1 and
in [68,69].
Figure 12: The moving base simulator at VTI. The projectors can be seen above the bodywork,
projecting a 120° visual field in front of the driver. The major components of the movement
system can be seen: the lateral sled, the pitch cradle and the roll axle below the projectors.
[Photo: Staffan Gustavsson (Redakta). Illustration: ARIOM.]
The surroundings are simulated and displayed to the driver via three main screens and
three rear view mirrors. Under the cab is a vibration table simulating contact with the
road surface, providing a realistic driving experience. This is also enhanced by realistic
environmental sound and light. In the 3-DOF moving base, the linear lateral movement
is large enough to allow realistic acceleration levels on a straight road; a total lateral
movement of 7 m is possible. It also includes roll and pitch movement that is used not
only for vehicle movement, but also for longer cornering situations and longitudinal
accelerations respectively. The vibration platform has possible motions in the vertical,
longitudinal, pitch and roll directions to simulate road irregularities (for a technical
specification see Table 2).
35
Section 4.3 Driving simulator tests
Table 2: Technical specification of VTI Simulator III [66].
Motion system
Pitch angle
- 9° to + 14°
Roll angle
± 24°
External linear motion
Maximum amplitude
± 3.75 m
Maximum speed
± 4.0 m/s
Maximum acceleration
± 0.8 g
Vibration table
Vertical movement
± 6.0 cm
Longitudinal movement ±6.0 cm
Roll angle
±6°
Pitch angle
± 3°
The visual system is optimized for short delay times to be acceptable for research use; a
maximum transport delay of 50 ms is specified. The visual system includes three video
channels projected at an angle of 30° by a 120° continuous screen ahead of the
bodywork. The resolution of the system is 1024 by 3840 pixels (i.e. 1024 by 1280 per
channel). The steering system has steering wheel torque feedback via an electric motor
controlled by a system with an update frequency of 200 Hz.
4.3.2 Curved cone track scenario
The aim of the curved cone track scenario, presented in Paper A, was to investigate
driver behaviour with a focus on the variation of different drivers’ ability to steer the
vehicle, i.e. path tracking skill. The moving base driving simulator was used to study
the drivers’ performance and behaviour when following a cone track scenario (see
Table 3 and Figure 13), primarily investigating the driven paths and lateral acceleration
levels. Relatively low speed was used, 55 km/h, creating a situation similar to some
road work conditions. The narrow cone track requires large lateral displacements and
relatively high attention being devoted to controlling the vehicle, but allows a limited
amount of variance in the chosen path. To investigate if there are differences between
drivers with a varying experience of driving, the recruitment base was chosen to be low
and high mileage drivers. All the drivers drove the simulator with the validated vehicle
model with both the standard setting as a reference and eight other combinations of
steering wheel gear ratio and steering wheel effort for comparison.
Table 3: Cone coordinates for the curved cone track scenario.
Left
[m]
x 75
107
139.1
151.25 184
216
248.75 280.9
313
-2.5
-1.5
5
6
6
5
-1.5
-2.5
Right
[m]
x 75,
107
139.75 151.9
184
216
248.1
280.15 313
y -6
-6
5
2.5
2.5
1.5
-5
y
-2.5
1.5
-6
36
Chapter 4 Measurement of driver characteristics
Figure 13: Simplified description of driving path for the curved cone track scenario. The full
width of the road can be used for the test, since the road shoulders have the same properties as the
area within the boundary lines.
4.3.3 Avoidance manoeuvre scenario
The objective of the avoidance manoeuvre scenario, presented in Paper B, was to
evaluate the relation between the driver skill and a large number of objective vehicle
parameters, all measured in the moving base simulator used. The recruitment of test
subjects was carried out based on a developed self-evaluation made by the drivers with
reference to descriptions of high skill and low skill drivers. A double lane change
(DLC) scenario specified according to ISO 3888-1:1999 was used both for recruitment
verification and the measurement of objective parameters. In Paper B, a suggested
method used for comparison of parameters under equal conditions is described. This
method was also used in Paper C and Paper E, and the results from Paper B are also
used in the development of the driver model in Paper D.
4.3.4 Driver response scenario
The line jump scenario, presented in Paper C, was designed and evaluated for the
investigation of driver-vehicle characteristics when following a movable reference line,
as seen in the illustration in Figure 14. The objective of this research work was to
investigate driver steering response during sudden unexpected reference path
movement, i.e. doing path tracking without preview, and this is possible with the line
jump scenario. The scenario design allows path tracking to be evaluated without the
uncertainty of a reference path coupled to a driver-preferred lane position or distance to
objects. With this scenario, the results instead show the effect of test subjects who can
choose how to perform the manoeuvre when guided precisely concerning time and
lateral distance, but not forced longitudinally by boundaries of a path, in a way very
similar to a regular step response.
37
Section 4.3 Driving simulator tests
+3.5
+2.0
+1.0
-1.0
-2.0
-3.5
Figure 14: Illustration of a driver (triangle) in a vehicle following a line movement +2 m. All the
different line movements used in the investigation are presented on the right hand side.
4.3.5 Curving road scenario
The objective of the curving road scenario, presented in Paper E, was to evaluate
objective parameters for characterization of driver skill when a driver is driving a
vehicle on a regular curving road (illustration shown in Figure 15). The reason for
performing this experiment is that the identification of driving skill in normal driving
conditions can be of great use for the setup of adjustable vehicle systems. A curving
road scenario was designed using both clear sight combined with high and medium
friction, and high friction combined with a limited sight distance. Fog was used to
investigate the effect of forced limitation of the driver preview distance. The method
developed compares parameters under equal conditions, identifying the ones with best
separation between the two recruited driver types.
1200
Global y-coordinate [m]
1000
800
600
400
200
0
-200
0
500
1000
1500
Global x-coordinate [m]
2000
Figure 15: Example sequence of road segments with an initial 300 m straight segment for
acceleration. The curves are highlighted in the figure with visualization of the curve radius.
38
Chapter 4 Measurement of driver characteristics
4.4 Proposed driving skill characterization
methodology
The proposed methodology, presented in Paper B, was used for comparison of
measures under equal conditions to identify parameters that are specifically useful for
characterizing the drivers’ path tracking skill. The parameters can be used for driver
metrics, representing typical driver characteristics in a driver model such as the one
presented in Paper D, for example. A method for pre-estimation of driver guidance and
control skill, also presented in Paper B, was developed to ensure that two
representative sets of drivers were recruited. These two sets are not homogeneous since
different driver behaviour to a large extent is shaped through individual experience. The
analysis was therefore performed using the two-thirds of the drivers (in the normal
case) which were the most separated for each specific parameter measure, calculating a
normalized value, the grade, which is used for comparison of the parameters:

if HVS > SS & LVS < SS
2
2
1 − C − abs( HVS − LVS ) ,
Grade = 
2 ⋅ SS
or HVS < SS & LVS > SS
2
2

0.5,
else
(7)
SS = Size of sections used in analysis
HVS = Number of high (or low) skill test subjects in the high value section
LVS = Number of high (or low) skill test subjects in the low value section
Te recorded data from these sections were used to derive driver metrics intervals, with
the higher graded characterizing the high skill and low skill differently.
4.5 Results from driving characterization
In the first driving simulator experiment, presented in Sub-section 4.2.3 and Paper A,
the drivers were only recruited with a rough definition of driver skill. However, the
results revealed some of the complexities in characterising and modelling drivers by
showing that the driver-preferred path strategy can differ a great deal between drivers
even for a relatively narrow track. Even though the vehicle steering wheel ratio and
effort were changed radically (but within realistic values), the driven path remained
relatively constant for the individual driver. Several drivers were easily separated from
each other in all of the approximately 30 to 50 runs. Some strategies resulted in twice
the lateral acceleration for some drivers compared to others, which also means a smaller
buffer to the limit set by road friction. This is important to take into consideration when
analyzing path tracking skill on a road or through a cone track, since the actual path
which the driver is trying to follow may differ in many ways from the optimal path, and
thus the drivers may be striving to fulfil different goals. However, as long as the main
goal is fulfilled, and no other measurable objectives have been given to the driver (such
as keeping a low lateral acceleration), it is not recommendable to make an evaluation
39
Section 4.5 Results from driving characterization
based on these aspects alone. Some experienced drivers tend to prefer to keep the
vehicle closer to the limit to be able to feel the transition towards tyre saturation better,
while others use their experience to drive with as large a buffer as possible. Examples
of the mean path over all the driver runs are shown in Figure 16, showing both test
subjects with very different paths (Figure 16a) and drivers with similar paths (Figure
16b).
a)
Test Subject
6
2
Distance [m]
Distance [m]
4
0
-2
-4
-6
b)
10
12
14
18
6
Test Subject
4
16
18
2
0
-2
-4
0
50
100
150
200
Distance [m]
250
300
-6
0
50
100
150
200
Distance [m]
250
300
Figure 16: Mean paths for test subjects with: a) large driving path difference and; b) small
driving path difference.
The cornering strategy is found to be one good measure for analysing the drivers’
decision making, while the magnitude of the standard deviation from the average path
is useful for analysing the repeatability of the drivers, and thereby provides an
indication of the precision with which the driver controls the vehicle. The ability to
handle system changes reflects how well the driver adapts to a new system.
The second driving simulator experiment included several different scenarios, presented
in Sub-sections 4.2.4-4.2.6, and in Papers B, C and E. This allowed the same drivers to
be used for all the scenarios (15 high skill drivers and 15 low skill drivers), increasing
the reliability of comparisons between the scenarios. In Paper B the avoidance
manoeuvre scenario is presented. This scenario was used both for validation of the
developed recruitment method based on self evaluation and to determine skilldependent metrics. The questionnaire-based recruitment process successfully provided
two distributions of drivers in the DLC-test, with only a very small overlapping of the
distributions of drivers with pre-estimated high and low skill for guidance and control,
as seen in Figure 17. The two driver populations are also shown to be significantly
different (p=0.000005).
Chapter 4 Measurement of driver characteristics
40
Number of Test Subjects
7
6
5
4
3
2
1
0
50
60
70
80
90
100
110
120
Velocity [km/h]
Figure 17: The highest velocities in a successful run in the simulator DLC-test, for drivers
recruited as being from either the low (white bars) or the high (grey bars) path tracking driving
skill population.
The position in the lane, y, has proved to be a parameter which re-occurs as high graded
for all the cases tested for this in the DLC simulation, i.e. 50, 60 and 70 km/h (all tests
subjects did not manage to drive at higher speed). The maximum value for this
parameter is constantly high graded, which indicates that the path strategy shows
similarities within the two groups of drivers. For the second half of the DLC-scenario in
this test, at the highest speed (70 km/h), several other high grade parameters are also
identified, including parameters related to both vehicle movement and steering wheel
movement. Since it is mainly the higher derivatives that qualify in this particular
scenario, this can be a problem for in-vehicle measurement due to the sensitivity to
noise. The standard deviation, however, which is calculated using all the data points
within the selected section, is relatively robust as a measure and is effective as a sorting
criterion for a number of parameters. For example, good performance can be possible
with reliable steering wheel angle measurement, since the standard deviation of both the
steering wheel angular rate and the steering wheel angular acceleration is shown to
exhibit relatively high grades.
In Paper C the line jump scenario, i.e. path tracking without preview, is presented with
results for all the drivers. It is shown that the overshoot amplitude increases for all the
cases when the velocity is increased, in most cases considerably. Another observation is
that, even though the overshoot increases with larger jumps in most cases, the
increment is far from proportional to the increment in the size of the jump. A short rise
time is shown to be representative of the high skill drivers in general, while a longer rise
time is shown to be representative of the low skill drivers, but, perhaps even more
importantly, this is also valid in combination with overshoot. The main reason for this
is likely to be that the high skill drivers have better knowledge of the vehicle response,
which is crucial for quick and accurate manoeuvres. Several objective vehicle measures
are shown to be important in characterizing the difference between high skill and low
skill drivers in straight line path tracking. The standard deviation of the lateral velocity
and yaw angle show particularly good separation of the two driver categories. These
measures relate to the stability of the vehicle after the manoeuvre. Values close to zero
are characteristic of the high skill drivers, with the standard deviation of the yaw angle
being lower than 0.15 degrees for all the tested jumps.
41
Section 4.5 Results from driving characterization
The results are useful in driver model development and driver skill identification
algorithms. Examples of other suitable usage of the simple principle of precision line
tracking are investigations of driver response to other forms of disturbances, e.g. side
wind or tyre punctures. Driving along a curved line in combination with response to a
variation of road friction is also a potential application.
Paper E presents a curving road scenario with the results of the driver characterization
method applied. The first two scenario sub-parts, called A-1 and A-2, are used with
clear sight with high road friction and medium road friction respectively. The remaining
five, called B-1 to B-5, are used with a decreasing sight distance (255, 135, 75, 45 and
30 m). Figure 18 shows all the scenario results for the smallest curve radius. Only the
centre parts of the curves are used, so that some of the effects occurring when going
from straight segments to curves are removed. Because of that, the shortest curve
lengths (see Figure 18b) have a sample length part of only ~40 m remaining for
analysis, making the analysis of these curves less reliable than that of the longer curves.
Driving in these shorter curves is also less similar to steady-state driving, since corner
cutting by the drivers is more likely to occur.
a)
b)
10
10
0.93
0.75
0.70
0.85
0.80
0.65
8
6
6
4
4
2
0
0.88
0.85
0.88
0.85
0.80
0.85
0.75
0.65
0.85
0.70
0.75
0.85
0.75
0.90
-2
-4
-6
0.73
0.85
0.90
0.65
0.65
0.70
High skill group
All
v 2x/R
2
0
0.50
0.50
0.50
0.50
0.50
0.60
0.60
0.60
0.50
0.80
0.65
0.60
0.70
0.85
0.75
0.75
0.85
0.85
0.90
0.60
0.60
3-1
A-1
3-2
A-2
4-1
4-2
4-3
B-2
B-3
B-1
Scenario part
4-4
B-4
4-5
B-5
-2
-4
-6
-8
-8
0.88
-10
Low skill group
0.73
8
Lateral acceleration [m/s2]
Lateral acceleration [m/s2]
0.93
3-1
A-1
0.88
3-2
A-2
0.90
0.75
0.80
4-1
4-2
4-3
B-2
B-3
B-1
Scenario part
0.90
4-4
B-4
0.80
4-5
B-5
-10
Figure 18: Lateral acceleration levels (minimum to maximum), with green for the low skill
section, red for the high skill section and black for all the drivers, in curves with a radius of 150
m. The grade is printed close to each bar. The dashed line shows the ideal lateral acceleration (ay)
according to the approximation ay=vx2/R. a) Left and right turn for a road turning angle of π/3
rad. b) Left and right turn for a road turning angle of π/5 rad.
For the fog-induced preview limitation, shown Figure 19, the number of measures with
high grades is starting to increase when the sight distance is shortened. This seems
reasonable, since drivers with low skill are expected to be more disturbed by this than
the skilled that should have more experience of difficult conditions. For extremely short
preview however, which is very difficult for all types of drivers, the number of high
grades decreases abruptly.
Chapter 4 Measurement of driver characteristics
42
Scenario A-1 measures
Additional measures
Poly. (Additional measures)
Measures with grade above limit
40
35
30
25
20
15
10
5
0
Scenario A-1
Scenario B-1
(255 m)
Scenario B-2
(135 m)
Scenario B-3
(75 m)
Scenario B-4
(45 m)
Scenario B-5
(30 m)
Figure 19: Number of high grade parameter measures in scenario B, divided into measures also
high graded in scenario A-1 and additional measures.
A comparison of the identified measures for the curving road with and without foginduced preview limitation is presented in Paper E, but it is also interesting to compare
these results to those for other scenarios. It is relevant to compare the line jump scenario
in Paper C with the curves in Paper D, since driving in both the line jump scenario and
the curving road scenario represents a quasi-steady-state behaviour in a relatively active
driving situation. The comparison shows good results for finding common parameter
measures that are graded high for both scenarios. Therefore, the results for these two
scenarios are presented in Table 4 with numbers indicating whether the measure is high
graded for both scenarios (blue highlighting), one scenario (grey for the curving road
scenario A-1 and hashed red for the line jump scenario), or none of the scenarios (no
highlighting). The yaw angle is not the same for the global reference and the road
reference, since scenario A-1 uses a curving road and the line jump scenario uses a
straight road (thus the hashed red for this parameter means that it is actually identical to
the road-referenced parameter, i.e. not new), but many other parameter measures are
found to be common for both the scenarios. The grade limit is set a little bit differently
in the two scenarios to create a balanced set with the fidelity possible.
43
Section 4.5 Results from driving characterization
Table 4: Number of occurrences of selected measures for scenario A-1 (grade equal or higher
than 0.78) in the in blue and grey and additional measures from the line jump scenario (grade
equal or higher than 0.75) hashed in red.
Measure
MAX L MIN L
MIN R MAX R
ABS
MEAN
STD
BW
POS
yroad
0
0
0
0
1
1
0
dyroad/dt
0
2
1
0
2
2
0
2
0
2
0
2
2
0
Parameter
2
d yroad/dt
2
Ψroad
0
2
1
0
2
2
0
dΨroad/dt
2
0
2
0
1
1
0
d Ψroad/dt
0
0
0
0
0
0
0
ay
2
1
2
0
2
2
0
day/dt
0
0
0
0
0
0
0
d ay/dt
Ψ
0
0
0
0
0
0
0
0
1
1
0
1
1
0
dΨ/dt
2
0
2
0
1
1
0
d Ψ/dt
0
0
0
0
0
0
0
β
0
1
1
1
1
1
0
dβ/dt
0
0
0
0
0
0
0
0
0
0
0
0
0
0
2
2
2
2
2
2
2
2
d β/dt
δSW
1
1
1
0
2
1
0
dδSW/dt
0
0
0
0
0
0
0
d δSW/dt
0
0
0
0
0
0
0
MSW
0
1
1
0
1
1
0
dMSW/dt
0
0
0
0
0
0
0
0
0
0
0
0
0
0
2
2
2
2
d M SW/dt
Since the parameters measures highlighted in blue in Table 4 have high grades both for
the driver response in the line jump scenario and the regular driving in the curving road
scenario, these show good potential for general usage. Without knowledge of the
curvature of the road, the standard deviation of both the steering wheel angle and the
lateral acceleration qualify for this. These can also be relatively easy to measure and are
not as sensitive as peak values for noise. If road curvature is known, the standard
deviation of both the lateral velocity and acceleration relative to the road and yaw angle
relative to the road could be used, given that the noise is low enough. With knowledge
of the actual levels of the parameter measures for the driver categories in the specific
driving case, there are several more parameters that could be interesting, e.g. the yaw
angle relative to the road, which is directly related to path planning and tracking.
44
Chapter 4 Measurement of driver characteristics
Chapter 5
Modelling of vehicles
This chapter gives a brief description of the two different vehicle models used: the
complex MBS-model in CarSim that is used as the controlled vehicle; and the simple
vehicle model that is used in the driver model as the description of the driver
understanding of the controlled system.
5.1 MBS-model description
The vehicle model used in the moving base simulator is a four-wheeled, 42-degree of
freedom (DOF) CarSim model, see Table 5 [68,69]. The sprung mass of the vehicle is
represented by a rigid body with six DOF. The front suspension and rear suspensions
are independent suspensions which has compliance in the lateral and longitudinal
directions. The longitudinal and lateral movements are constrained as functions of
vertical movement described by nonlinear tables, the camber and toe angles related to
the vertical position by nonlinear tables, and the suspension springs include hysteresis
due to friction. The dampers produce forces as nonlinear functions of the stroke rate.
The suspension roll moments include a nonlinear auxiliary roll moment to account for
roll stiffness beyond the effects predicted by the spring properties and geometry. Each
wheel has compliance that affects the toe and camber in response to the tyres’ shear
forces and moments. The roll and jacking forces are calculated as the natural results of
full 3D kinematical curves and compliance effects. Each wheel has one spin DOF and
each tyre has two dynamic DOF: one for lagged lateral slip, the other for lagged
longitudinal slip. The main model uses nonlinear tables to represent the lateral force,
longitudinal force, aligning moment, and overturning moment as functions of slip, load,
and camber. The lateral and longitudinal forces and moments are combined using
combined slip theory as described by Pacejka and Sharp [70].
45
Chapter 5 Modelling of vehicles
46
Table 5: Degrees of freedom in the CarSim vehicle model [68,69].
Sprung mass
Vertical movement of front suspension
Vertical movement of rear suspension
Front suspension compliance
Rear suspension compliance
Wheel spin
Lagged lateral slip
Lagged longitudinal slip
Friction in suspensions
Friction in tyres
6 DOF
2 DOF
2 DOF
6 DOF
6 DOF
4 DOF
4 DOF
4 DOF
4 DOF
4 DOF
The equations of motion are derived from the first principles for fully nonlinear 3D
motions of connected rigid bodies. The equations of motion are ordinary differential
equations (ODEs) and can be solved with most numerical integration methods. The
ODEs are solved using a second-order Runge-Kutta algorithm [69]. The vehicle model
used represents a medium-sized car. An external servo is added to the vehicle model for
better representation of the steering system in the calculation of the torque at the
steering wheel.
5.2 Validation of MBS-model
When a moving base vehicle simulator is used for behavioural research, it is of the
utmost importance that the movement felt by the drivers is realistic. This does not
necessarily mean that the model has to resemble an existing vehicle, but it should
behave as a vehicle could be expected to behave in order to be predictable by the driver.
It is of course important that the vehicle model should be validated against real vehicle
dynamics if the results are to be applicable for the specific vehicle, but this is not as
important here as the requirement that driving the vehicle model should feel the same as
driving the vehicle modelled. This feeling is dependent on both the vehicle model and
the simulator. The CarSim vehicle model used in the simulator tests performed here
has been validated by Saab Automobile within the project. Examples of validation
results are shown in Figure 20.
47
Section 5.3 Driver’s internal vehicle model
a)
b)
40
10
20
Yaw Rate [deg/s]
Ay [g]
5
0
0
-20
-5
-40
-10
1
2
3
4
5
6
-60
1
2
Time [s]
3
4
5
6
Time [s]
Figure 20: Examples of validation results using 0.7 Hz sinusoidal input with increased amplitude
(dashed line for simulations and solid line for measurements, with the same colour for
corresponding test cases). a) Lateral acceleration as function of time. b) Yaw rate as function of
time.
The vehicle model has also been verified in the simulator to be acceptable by several
drivers, of which one is a test driver who is familiar with the specific modelled vehicle.
The overall behaviour is considered very realistic for lateral movement, but in some
special situations with very high lateral acceleration, and in some situation with
transition from a straight road to long curves, there are some unresolved discrepancies
in movement that can be noticeable (at least by a trained driver). With the limited
possibility of generating force for longitudinal dynamics in the setup used for the
simulator, i.e. only by tilting the cabin, it is not possible to make this movement with
enough realism here, and therefore the acceleration and deceleration phases are kept to
a minimum and are not used for analysis.
5.3 Driver’s internal vehicle model
This section describes the fundamental lateral vehicle dynamics that, with appropriate
parameter values, can be used for representation of the driver’s knowledge of the
controlled vehicle. Where parameter and variable descriptions are left out, these are
found described in the Nomenclature. Knowledge of the driven vehicle is a part of
driver modelling that can include a dynamic model or more static information. The
driver model presented in Paper D is based on the fundamental dynamics presented
here, using a quasi-static approximation instead of the full dynamics.
The vehicle equilibrium equations in Equation 1 describe the movement of a frontsteered vehicle approximated with two wheels, no height, and no influence of external
forces. They are also the basic equations for the simplified vehicle model often referred
to as the bicycle model or one-track model. This is a basic vehicle description that is
used as a basis for many driver models, and it is employed here as well, since it offers a
good approximation of the controlled system. If a more complex description using, for
example, four-wheel steering is sought, this can be found in [71] by Abe, for example.
Chapter 5 Modelling of vehicles
48
m(v& x − ψ& v y ) = − F12 sin δ
m(v& y + ψ& v x ) = F34 + F12 cos δ
(1)
J zψ&& = f F12 cos δ − b F34
For a more commonly used form of the bicycle model, the last two rows in Equation 1
are rewritten to the matrix form in Equation 2. In this form the tyre forces are
approximated with a linear relation to the slip angle of the wheels, and also small angle
approximations are used for the tyre slip angles and the steering angle at the wheel.
Since both the slip angles and the steering angles are rarely much larger than 5°, the
error will be kept relatively small (in the order of magnitude of 1%) for the driver
scenarios in this research work.
C + C34

 mD + 12
vx

 f C12 − bC34

vx

f C12 − bC34  v y   C12 
  

vx
δ
  = 

f 2 C12 + b 2 C34   
JzD +
  

vx
 ψ&   f C12 
mv x +
(2)
The equations presented in Equation 2 can in the steady state, i.e. v& y = ψ&& = 0 , be
expressed as:
δ = ψ& ⋅
L2 C12 C34 + mv x2 (bC34 − f C12 )
=
v x LC12 C34
(3)
m(bC34 − f C12 )
L
L
= ψ& +
v xψ& = ψ& + K us v xψ&
vx
LC12 C34
vx
where the vehicle understeer characteristics are described by K us , which is constant if
the axle characteristics are approximated to be constant:
K us =
m(bC 34 − f C12 )
LC12 C 34
(4)
With zero body slip angle, the velocity body fixed coordinates (vx,vy) will be equal to
the velocity in the rotating reference frame (vT,vR), with one tangential component and
one radial component, directly related to the circle defined by the turning radius. With a
body slip angle different from zero, however, there will be a difference in velocity
between these systems. This difference can be calculated with:
49
Section 5.3 Driver’s internal vehicle model
 vy
 vx
β = arctan
vT2 = v x2 + v y2
⇒ v x = vT


 ⇒ v y = v x tan β 
2
2
2

 ⇒ vT = v x (1 + tan β ) ⇒


1
(5)
(1 + tan 2 β )
With values inserted this shows, for example, that a slip angle of 8° would result in less
than 1% error for the approximation (1 + tan 2 β ) ≈ 1 , i.e. v x ≈ vT . This is small enough
to allow the small angle approximation here as well, and Equation 3 can be rewritten
according to:
δ=
L
L
L
ψ& + K us v xψ& ≈ ψ& + K us vTψ& = + K us a R
vx
vT
R
(6)
The term L/R can be referred back to a case with lateral acceleration close to zero,
meaning that the forces generated by the tyres are small. In such a case the geometric
steering has much higher relevance than the tyre characteristics or other dynamic
properties of the vehicle. This baseline geometric steering wheel angle needed for the
specific turn radius is often referred to as the Ackermann angle [72,73], which for the
bicycle model is simply δ A = L / R , as presented in Figure 21. Geometrically the angle
would be δ A = arctan(L / R ) according to the figure, but, as for the other small angle
approximations used, for L much smaller than R the simplified (and commonly used)
description will differ very little from the more complex one.
δA
L
δA
R
Figure 21: Illustration of the Ackermann angle (δA) for a bicycle model.
The term Kus in Equation 6 is a linear approximation that assumes linear dependence of
the slip angle and the lateral force at the wheels, which usually is an acceptable
simplification for small slip angles, but which does not hold true for large angles, where
50
Chapter 5 Modelling of vehicles
the lateral force is saturated. The lateral acceleration ay can also be used instead of the
radial centripetal acceleration aR with the small angle acceleration, since the difference
will be only cosine (β), which for a reasonably large slip angle of 8° is less than 1% as
well.
Chapter 6
Modelling of drivers
This chapter gives a brief overview of different driver model types: the compensation
tracking type, which does not include any future path information; the preview tracking
type, which does take the basic human ability to utilize preview into account; and the
fuzzy type, using a blend of simultaneous feedback processes.
6.1 Compensation tracking models
A basic block diagram illustrating the compensation tracking models is shown in Figure
22, where G(s) represents the transfer function of the vehicle system and H(s)
represents the transfer function of the driver compensation [74]. Where parameter and
variable descriptions are left out, these are found described in the Nomenclature.
ε
r
H(s)
δSW
G(s)
y
-
Driver model
Figure 22: Basic structure of compensation tracking models [72].
The model presented by Iguchi in 1959 used Equation 8 as a model for the driver block
(Figure 22) [75,76,74].
H (s) =
Kd s 2 + K p s + Ki
(8)
s
51
Chapter 6 Modelling of drivers
52
The model is simple, but the coefficients K d , K p and K i are difficult to determine and
therefore the model has not been used extensively [74].
Ashkens and McRuer presented the model in Equation 9 for the driver block in 1959,
introducing the vehicle-independent driver brain response delay, td , and the driver
action delay, th . Lead, TL , lag, Tl , and gain, K , however, are dependent on the driver
experience and the vehicle driven [77, 74].
H (s) =
Ke − td s (1 + TL s )
(1 + t h s )(1 + Tl s )
(9)
In 1967, McRuer presented a method for determining TL , Tl , and K by fitting
Equation 10, where ωc is the cross-over frequency of the open-loop function,
H ( s)G ( s) , often referred to as the cross-over model [13,16].
H ( s)G (s ) =
ωc e −t
ds
(10)
s
These models can be useful for corrections of identified errors in specified conditions,
but are not fully satisfying for more complex situations due to the simple feedback and
lack of effective preview usage.
6.2 Preview tracking models
In the preview tracking models, a preview strategy element and feedback function are
included. A basic block diagram illustrating the preview tracking models is shown in
Figure 23, where P(s) represents the preview strategy, H(s) the control characteristics
function, and B(S) the driver’s feedback (or prediction) function of vehicle motion
giving the driver estimation of future lateral position after driver compensation, yp,
aimed at matching the output from the preview strategy, f0 [74]. Where parameter and
variable descriptions are left out, these are found described in the Nomenclature.
r
P(s)
ε
f0
δSW
H(s)
-
yp
B(s)
Driver model
Figure 23: Basic structure of preview tracking models [74].
G(s)
{y}
53
Section 6.2 Preview tracking models
The first preview tracking model, the linear prediction model, shown simplified in
Equations 11 to 13, was presented by Kondo in 1968 [78,74]. This model did not take
the response delay of the driver into account, but introduced the predicted lateral
position, yp, at time t + T p based on the lateral position, y, and heading angle,ψ , at time
t, together with the vehicle speed, V [74].
T s
 P(s) = e p

H (s) = K
 B ( s ) = (1, T V )
p

(11)
y (t ) = [ y (t ),ψ (t )]T
(12)
y p (t + T p ) = y (t ) + T pVψ (t )
(13)
By replacing Vψ (t ) with y& (t ) in the equation above for small ψ (t ) and low frequencies,
a single variable feedback can be formed, B( s ) = 1 + T p s [79,80,81,82].
In 1968, Yoshimoto presented the second order prediction model, where both a second
order prediction feedback and a driver response delay are included, as well as an
integration block with an empirically determined K to represent the correction ability of
drivers [83,84,85,74].

 P ( s ) = e Tp s

K −t s

H (s) = e d
s

2
Tp 2

 B ( s ) = 1 + T p s + 2 s
(14)
In 1966, Sheridan was the first to present the optimal control concept, in which he
considered the driver/vehicle tracking problem using a local optimal preview model
where the driver aims at minimizing the tracking error looking over a finite interval of
the future path [86,74]. MacAdam, however, developed the concept further in 1980, and
with his work the optimal preview control model was taken from theory to engineering
application [25,26,74,82].
An extension of the optimal preview control model is presented by MacAdam in [62],
using a nonlinear vehicle description. The driver model was developed as a GM project
to cover “near/at-limit vehicle handling”, with “substantial utilization of available
tyre/road friction”. A key factor in the new model was the inclusion of driver
characteristics, e.g. look-ahead/preview sight information; the ability to adapt vehicle
dynamic properties at varying adaptation rates; compensatory abilities to alter preview
utilization; and anticipatory abilities based on upcoming road geometry or path
requirements; as well as the inclusion of limitations, e.g. reaction time delays; neuromuscular dynamic lag; and corresponding frequency response characteristics. In [27],
Chapter 6 Modelling of drivers
54
which focuses on the control aspect of human driving with an internal model in a way
similar to that in [62], MacAdam lists what he considers to be a minimal representation
of the human driver:
•
Transport time delay
•
Preview for upcoming lateral and longitudinal requirements
•
Driver adaptation provision
•
Cross-over model behaviour near the cross-over frequency
•
Internal vehicle model for prediction
He also lists additional and desirable features:
•
Neural delays, thresholds, rate-limiting, and dynamic properties of individual
sensory channels
•
Neuromuscular filtering
•
Previewed path adjustment capabilities and strategies to adjust for e.g. skill
and style
•
Speed adjustment based on upcoming lateral path requirements
•
Surprise or situational awareness features
The situational awareness feature is modelled in the GM-UMTRI model [62] as a
cognitive/recognition delay to simulate the “casual” driver control behaviour. What this
does is to delay the desired path information until a certain lateral acceleration limit is
exceeded, i.e. a conflict or sudden manoeuvring condition is occurring. This function is
then deactivated for the remaining part of the simulation run. An illustration of the GMUMTRI model can be seen in Figure 24, followed by a short description of the ten
elements.
55
Section 6.2 Preview tracking models
DRIVER MODEL BLOCK
EXTERNAL VEHICLE
SIMULATION PROGRAM
1
PREVIEW SCENE
present vehicle
response
2
Driving Simulator Test 1
SENSORY
LIMITATIONS
& NOISE
3
7
PATH
PLANNING
10
INTERNAL
VEHICLE DYNAMICS
8
”situational
awareness
response
time(s)”
T
PREVIEW PATH
OBSERVATION
CAPABILITY
4
PREDICTION CAPABILITY
9
predicted
path
5
STEERING CONTROL
CALCULATION
SPEED
CONTROL
desired path
6
optimal steering
control
DRIVER PHYSIOLOGICAL
& ERGONOMIC CONSTRAINTS
DRIVER MODEL STEER CONTROL
SPEED ADJUSTMENT REQUEST
Figure 24: GM-UMTRI driver model with reference numbers for each element [62].
The components of the driver model are listed below according to the reference
numbers in Figure 24:
1.
Previewed scene (desired path or road input description)
2.
Sensory limitations and noise (pertaining to incoming vehicle response
signals)
3.
Internal vehicle dynamics (4 degrees of freedom)
4.
Prediction capability
5.
Steering control calculation
6.
Driver physiological and ergonomic constraints (associated with driver
steering and control responses)
7.
Path planning options (centre-line smoothing, or minimum curvature path)
8.
Fixed or variable driver preview capability (based on upcoming vehicleboundary constraints and projected interferences)
9.
Driver speed control (for accommodating upcoming lateral path requirements
and estimated lateral demands)
10. Situational awareness parameter (a simple delay that only affects the path
input channel and only during the initial portion of a manoeuvre)
56
Chapter 6 Modelling of drivers
MacAdam does not claim to cover more than the control activities of driving in the
described model, even though some activities may go beyond low-level controlling
(e.g. driver preview used for adjustment of the speed and track). The model also treats
longitudinal and lateral control behaviour separately in the current form, but the
combined form of control behaviour that more naturally represents the behaviour of
human drivers is described as appealing for future modelling efforts. MacAdam also
recognizes driver skill-related issues and techniques for recognizing and representing
driver skill as an area that requires more research work. Moreover, the areas of “smart
vehicles” potential interaction with human drivers, better understanding and
categorization of less skilled drivers, improvements in the understanding of how drivers
internalize their view of the external world/controlled vehicle, and the modelling of
this, are to be considered for further work, according to MacAdam [62,27].
6.3 Fuzzy set theory models
When mathematical models are unavailable or too complex for the required accuracy
and speed of response, fuzzy set theory may be helpful. This allows objects to have
partial membership of a set, which is not the case for conventional (crisp) set theory.
Human perceptions are more naturally defined by fuzzy sets than by crisp ones, and
fuzzy mathematics combined with knowledge-based (expert) logic is what gives the
fuzzy logic system the ability to “reason” like humans. Fuzzy logic is, however, not as
easy to validate as conventional control theory and empirical methods may be the only
way to accomplish this. Fuzzy logic is considered best when a process can be defined
with IF-THEN rules, while neural networks are useful when only input-output signals
are known [87]. What neural networks really do is to classify data by matching signals
to learned patterns, i.e. they provide a pattern recognition method. Neural networks
have different properties, depending on how they learn and train. Some networks train
only on input data and are particularly good at spotting similarities. The dynamics of
these networks are sensitive to repetition, which allows them to evolve transfer
functions influenced by natural clustering in the data. Besides being adaptive, neural
networks are also robust, since they can give quite good answers even when the input
data are noisy or incomplete. Neural networks learn by reading known input/output
samples and adapting themselves to map them together, which means that they do not
need explicit equations to correlate this relation. This is described by MacAdam in [27]
not only as a benefit, since the lack of parameters directly linked to physical
characteristics of the driver often makes the results difficult to interpret. Berardinis
asserts in [87] that, even if the right network is used for a particular system, this is still
not a guarantee of success. If the training data is bad, the result will be bad. It is also
important to know that a network performs badly both with too little and with too much
training. Trivial relations in the training set which are of no interest for classification
may also be a cause of error. As stated by Neusser in [88], one of the major benefits of
neural networks is that they are capable of learning complex, highly non-linear relations
between input and output, even if these relations are not explicitly known; i.e. the
network will learn to extract the essential sensor information, and an actual neural
hardware realization will be fast, as neural networks are naturally suitable for a
massively parallel execution.
57
Section 6.3 Fuzzy set theory models
Neural networks are used by MacAdam in [55] for identifying and classifying the onhighway longitudinal control behaviour of drivers based on different levels of displayed
aggressiveness, and also for representing or modelling instances of longitudinal control
behaviour (Figure 25).
TIME HISTORY
CATEGORIES OF
INPUT PATTERNS
BEHAVIOUR
CLOSING-IN RAPIDLY
RANGE
RANGE-RATE
NEURAL NET
CLASSIFIER
OF
DRIVING
BEHAVIOUR
CLOSING-IN
FOLLOWING
FALLING-BEHIND
FALLING-BEHIND RAPIDLY
Figure 25: Neural net pattern recognition for classifying driving behaviour [55].
The tests were conducted under normal highway conditions in the USA with
measurements of the range and range-rate (with an infrared sensor), driver steering,
throttle control, yaw rate and lateral acceleration, for example. Driver style
classification was carried out based on the displayed level of aggressiveness in terms of
willingness to pass, follow, or be passed by other vehicles. The study shows that a
neural net representation derived from prior training data does not necessarily predict
future driver control behaviour under similar operating conditions accurately, which can
be seen in Figure 26, where the driver behaviour changes several times during driving.
35%
30%
25%
20%
15%
10%
5%
0%
CLOSING IN
RAPIDLY
CLOSING IN
FOLLOWING
FALLING
BEHIND
FALLING
BEHIND
RAPIDLY
Figure 26: Distribution of driving behaviour identified for an “average” driver over a one-hour
period with 135 events [55].
58
Chapter 6 Modelling of drivers
MacAdam suggests that this can be explained partly by the fact that the longitudinal
control task in the study was affected by components of the driving environments, such
as the road grade, nearby vehicles, visual distractions, and other influences that are not
reflected directly within the limited sensor information provided only by the range and
range-rate measurements in the study. He also suggests that casual control activity
perhaps does not demand the same level and continuity of attention as is required for
path following. It was concluded that drivers are probably affected by other influences
beyond just the range and range-rate, which implies that a dynamic classification of a
driver is preferable.
In [89] Ohno uses a three-layer feed-forward neural network model with a sigmoid-type
activation function for the units in the hidden and output layers. This model performs
both control and learning at the same time. The study is focused on the use of adaptive
cruise control (ACC) and an important lesson is that some drivers tend to over-trust the
technology and put the driver-vehicle system in a situation where an accident cannot be
avoided.
Although neural networks and other similar pattern recognition methods that use
automatic training are very powerful for recognition of complex patterns, they are
sensitive to the training procedure, and since a mix of parameters are used, it is often
difficult to make any detailed conclusions of the reasons behind the results and the
internal process. With sufficient knowledge of the internal structure of the modelled
system, and a careful selection and usage of parameters, these methods can be very
useful for identification of system state and to model complex decision-making. For
lower level control and regulation tasks however, there are more benefits of using a
model which is more explicitly defined, and with an isomorphic model in which the
internal structure of the system is also described, it is possible to set parameters that are
relevant for the internal process.
Section 6.4 The KTH Vehicle Dynamics driver model
6.4 The KTH Vehicle Dynamics driver model
The objective of the KTH Vehicle Dynamics driver model is to create a model that can
be configured to represent a typical high skill and a typical low skill driver in a path
tracking scenario with constant driving speed. The model described in Paper D is a
preview tracking model with feedback of the lateral position of the vehicle’s centre of
gravity (CG) relative to the preview path and yaw angle with a curvature reference to
the path.
A driver model is proposed that is based on a relatively simple internal vehicle model.
The driver model is flexible and intuitive for the setting of physically relevant
parameters and the current design shows that both simulation of general driver
characteristics and differentiation between high and low skill driver behaviour can be
accomplished. By using the moving base driving simulator, VTI Simulator III
[65,66,67], integrated with the desktop vehicle simulation program CarSim [68], it is
possible to use the same validated vehicle model in both the driver model and the
simulator and only replace the source of input. The model separates three levels of
driving skill: perceptual, anticipatory & interpretational; decisional and; execution
skills, into different blocks.
The validation of the model is performed using the results from driving simulator tests
with the ISO 3888-1:1999 double lane change scenario. The parameter sets used for the
model configuration are selected based on physical relevance to the model, and
optimization is carried out with a Nelder-Mead implementation [90], which is based on
the Simplex method [91]. The driver model that is presented here has been shown to be
able to resemble the characteristics of different driver types in a path tracking scenario
for 70 km/h (examples are shown in Figure 27), and with reasonable modifications the
driver model can represent drivers at other speeds. Since the settings are composed of
driver type specifications for each measure for the groups of drivers, individual drivers
fulfil most of the metrics in the same run but not necessarily all of them, which is also
shown in Paper B.
59
Chapter 6 Modelling of drivers
60
a)
c)
50
δSW [deg]
100
50
δSW [deg]
100
0
-50
-100
0
20
40
60
80
distance [m]
100
120
b)
0
-50
-100
0
20
40
60
80
distance [m]
100
d)
6
6
Tracking path
y
5
y
road
Tracking path
y
5
road
4
road
+/- 0.89m
y
4
3
2
1
2
3
4
5
6
1
0
-1
+/- 0.89m
3
2
1
2
3
4
5
6
1
2
3
4
5
6
20
40
1
0
-1
-2
-3
road
δSW
y road [m] , δSW [deg/50]
δSW
y road [m] , δSW [deg/50]
120
0
1
2
20
40
3
4
60
80
distance [m]
5
100
6
120
-2
-3
0
60
80
distance [m]
100
120
Figure 27: Measured driving simulator data for steering wheel angle using sample drivers for
three consecutive runs (from the test in Paper B), and the driven path of the driver model using
the two skill settings derived from optimization at 70 km/h. The cones are marked with red and
green circles, the extrapolated width of the vehicle is presented in black (not yaw-compensated),
and the distance is measured in metres. Dotted vertical lines indicate estimated time stamps in
seconds. a) Actual runs from one high skill driver. b) Driver model results for the high skill
setting. c) Actual runs from one low skill driver. d) Driver model results for the low skill setting.
Chapter 7
Scientific contributions
This chapter lists the main scientific contributions of the thesis and appended papers:
1.
The identification of very different driver-selected paths in a narrow cone
scenario in Paper A. In addition, the identification of driver differences in
adaptation to different steering settings by comparing changes in the standard
deviation from the mean path.
2.
The recruitment process described in Paper B, enabling a good selection of
drivers with high and low path tracking skill prior to any actual driving tests.
3.
The design of the simulator tests presented in Paper C and Paper E,
providing a valuable base for analysis of the path tracking behaviour:
a.
The line jump scenario in Paper C, which removes the uncertainty of
driver path selection by using a single line to track, and the preview
strategy by instantaneous movement of the line.
b.
The curving road scenario in Paper E, which allows regular
countryside driving to be analysed for different curve types and
straight segments individually, which also allows straightforward
comparison of different situations, e.g. the varied preview limit used
in the paper.
4.
The method for evaluating the different objective parameter measures, the
grade-calculation presented in Paper B, which simplifies creating the metrics
with intervals that can represent drivers with typical path tracking
characteristics.
5.
The parameters identified as skill-related for the different scenarios, with
emphasis on the potential benefit of the curving road scenario if used in
identification of driver skill in real vehicles.
6.
The driver model in Paper D, using physical parameters and a simple internal
vehicle model to enable the setup of models describing different types of
driver skill, with validation against metrics gathered in a driving simulator test
using the same scenario.
61
62
Chapter 7 Scientific contributions
Chapter 8
Discussion and conclusions
The aim of this work was to determine the driving skill definition to use, define the
recruitment process, and develop a method to characterize the driving skill. Also, a
driver model was to be developed using metrics from objective measures related to the
driving skill.
In the first study, presented in Paper A, it is observed that different strategies can be
found for drivers in a relatively narrow cone track, which indicates that it should be
possible to observe some differences besides those concerning the actual hitting of
cones. However, with a small number of test subjects and a relatively rough definition
of driver categories, it is not obvious how to relate these differences to driver
experience or driving skill. In the following simulator tests a thorough investigation of
objective parameter measures was conducted using test subjects recruited as being
representative of high skill and low skill drivers. A recruitment procedure was proposed
and evaluated on 30 test subjects, and the verification of the recruited set of drivers in a
double lane change scenario shows a wide variation in the actual performance, but only
a small overlap between the recruited groups, thus verifying a successful recruitment of
test subjects for this test. Since the term skill is quite complex, it is also often difficult
to find a consensus as to what the term comprises and what should be measured.
However, this study is focused on finding metrics that can be used successfully for
describing the skill level of the group of drivers in some very different situations, not to
describe the behaviour of a specific driver in every situation. Below are some
comments about the use of the scenarios in this work:
•
For the double lane change scenario (Paper B), objective parameter measures
sampled in the second half of the double lane change manoeuvre for the higher
velocity are found to be most useful for categorization of drivers (with good
correlation to drivers recruited as low skill and high skill drivers), which can be
derived from the fact that inadequate driving skill becomes more evident in the
second part of the scenario. The standard deviation of the steering wheel rate
and the standard deviation of the angular acceleration are both measurements
that were found useful.
•
For driver skill characterization, the line jump scenario (Paper C) has been
analyzed not only with measures describing the response of the drivers
63
Chapter 8 Discussion and conclusions
64
primarily by their behaviour during the first few seconds after the line jump,
but also with the objective parameters used in the other scenarios, calculated
from data at the time from 4 to 8 s after the jumps. The last four seconds are
used for emphasis on the straight line path tracking skill after the large
movement from the previous position of the line (centre of the road).
Relatively good categorisation performance is found for some measures, both
for the first and last part. A short rise time, by itself and also in combination
with a small overshoot, is shown to be an important characteristic of the high
skill drivers. Several of the objective vehicle parameters also demonstrate a
difference between the groups of high skill and low skill drivers, with low
standard deviation of the lateral velocity and the yaw angle as examples that
can be interpreted as typical high skill driver characteristics.
•
For the curving road scenario (Paper E), the curves are found to be more
reliable for identifying driver skill than the straight road segments, and a
number of measures show good performance in characterizing driving skill
under the tested conditions, both for clear sight and with the preview limited
down to 30 m. The standard deviation proves to be very useful as a measure,
and qualifies for successful driver skill categorization for commonly sampled
data such as the lateral acceleration, yaw rate and steering wheel angle.
However, since these measurements are taken exclusively during curve
negotiation, these situations need to be identified during driving. Moreover,
even though the parameters that are readily measurable can be sufficient, a
significant improvement of the number of useful measurements can also be
achieved if accurate information about the road curvature is available. This
should be possible to solve using state-of-the-art GPS-systems. When the
preview shortens, new measures appear that also separate the driver groups,
but for an extremely short preview the number of separating parameter
measures is reduced rapidly, which can be an effect of both the high skill and
low skill drivers performing at an equally low level. These results should be a
good starting point for online classification of driving skill that can be used for
adaptation of adjustable vehicle systems to aid and support the driver with the
driver ability taken into consideration.
Since the parameter measures which can be used with most success for the
characterisation of driver path tracking skill are different for different scenarios and
scenario sections, and since drivers may be skilled at different tasks, the best results are
found when knowledge is possessed of such things as the curvature of the road (or at
least whether or not the road has a curvature at the present location of the vehicle) or
the type and part of the current manoeuvre. However, good correlation between results
and driving skill is found for several cases using common parameter measures, and the
case studied here involving normal driving conditions (a curving road) seems to be
excellent for categorizing drivers based on the skill level. The double lane change
manoeuvre and the line jump scenario have proven to be very interesting as well, since
they show relatively high driver classification validity for a number of parameter
measures. However, the double lane change scenario measures are not directly
comparable with the measures of the other scenarios, since the double lane change is
analyzed without a relevant path reference (due to the more apparent possibility of
selecting different path strategies within the cone track).
Section 6.4 The KTH Vehicle Dynamics driver model
Parameters measures with high grades both for the driver response in the line jump
scenario and the regular driving in the curving road scenario on the other hand show
good potential for general usage. Without knowledge of the curvature of the road, the
standard deviations of the steering wheel angle and of the lateral acceleration are such
measures. If the road curvature is known, the standard deviations of the lateral velocity
and of the acceleration relative to the road, and the standard deviation of the yaw angle
relative to the road, should be of interest as well.
The developed driver model presented in Paper D includes a quasi-static vehicle
description and is quite a simplified representation of a driver. On the other hand, it
enables straightforward modifications of physical parameters that affect the model
behaviour to represent both typical high skill and typical low skill drivers, and this can
be much more complicated in a dynamic model since it involves more parameters being
set. A complex model may also be too advanced to be a realistic representation of the
driver anticipation of the vehicle motion, while the vehicle understeer behaviour used is
quite an intuitive representation of the vehicle characteristics. Nonlinearities in a more
advanced model may also be difficult to use inverted, and therefore one may have to
rely instead on interpolation between the results of several tested steering angles as in
[62]. The vehicle model is a validated standard configuration of a medium-sized
passenger vehicle, and changing this model to another may require some adjustments of
the driver model settings. Although it was concluded in Paper A that drivers tend to
choose the same path to track, even when the steering characteristics of the vehicle are
changed, the vehicle characteristics may still affect the driver action as described in this
thesis. The ISO 3888-1:1999 double lane change scenario was chosen for the initial
setup of the driver model, since this manoeuvre is quite demanding and thus also
demanding for the driver model. Using the driver model in another scenario, e.g. for the
curving road in Paper E, requires that the parameters are identified for the driver
behaviour in this situation, since the requirements on the drivers are different, and this
can be accomplished with the same method as that presented in Paper D for the double
lane change.
To conclude, the developed methodology for finding objective parameters that are
typical of driving skill has proven successful. The results from the simulator
experiments show that several parameters have measures that can be used to describe
driver skill, and several have measures that are very useful for driver modelling. A
model has also been presented that can be set for the driver skill. If driver skill
characterization is used in the driven vehicle, this also gives a potential benefit for
automated adjustments of vehicle systems.
65
66
Chapter 8 Discussion and conclusions
Chapter 9
Recommendations for future
work
The driver skill characteristics identified here can be used both for the design of a driver
model and the classification of drivers, but only the driver model application has been
explored here. The measures with a high grade could be combined to provide a higher
possibility of correct classification. A good preview path is also an important building
block of a driver model, and although the paths used here are based on the results of
real drivers, a good method for creating artificial paths is to be preferred to improve the
flexibility of the model. The path used here shows remnants of the individual drivers,
and although a filtering process would smooth the path out, there could still be low
frequency parts that are unwanted. It is also seen in the results that the frequency of the
steering signal from the driver model in some cases features more periodic high
frequency contents than would be expected from a real driver, which may be avoided
by the addition of metrics directly linked to the behaviour in the frequency domain.
Moreover, a more advanced optimization routine that succeeds better in finding the
global optimum may also result in sets of parameter values that are even better at
representing the two driver categories. The identification of driver types in running
vehicles on any road should be explored further, since this unfortunately did not fit into
the project timeframe. Pattern recognition routines could be trained with data found
relevant to driving skill in given situations, for which the baseline is given here.
Even though it was concluded in Paper A that drivers tend to choose the same path to
track even when the steering characteristics of the vehicle are changed, depending on
which parameters are used, greater or smaller adjustments have to be made to reflect the
driver behaviour in a vehicle with new characteristics. It is obvious that the vehicle
representation in the driver model must be changed according to the new vehicle, but
changed vehicle characteristics may also affect the metrics used for the driver
categories, due to driver shortcomings in adapting to the new vehicle (resulting in
degradation in performance), due to the enabling of a higher (or lower) driver-vehicle
performance, or due to other vehicle response changes which the driver voluntarily or
involuntarily is not cancelling out through changes in his or her behaviour. Another
aspect is, of course, the use of different scenarios and velocities for which validation
67
68
Chapter 9 Recommendations for future work
should be performed as well if the model is to be used for such cases. The curving road
scenario does feature a variation in curve types, but this variation does not cover more
than two radii. Even though a general driver model or identification algorithm could be
configured to interpolate and extrapolate from the results, it is obvious that more tests
could be necessary, especially with driving on real roads. These roads are also more
irregular in shape than the currently tested artificial roads. The road used in the driving
simulator is flat and the constant radius curves are placed directly after the straight
segments, and this is not the most common configuration found on roads built today. To
make curves easier for the driver to negotiate, the Swedish design rules suggest that
sharp curves should begin with either a transition curve with half the curve radius or
with a clothoid [92]. Moreover, banking of curves is a measure that is often used, and
this can be accomplished with different centres of rotation [93]. Therefore, it is
suggested that real roads or simulated roads with more realistic features should be
examined to gain further knowledge of the effect of these variations.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
Willet, T.C. Criminal on the road, Travistock Publications Limited, Great Britain, 1964.
A brief history of registration, Vehicle Policy Group, SV2, Driver and Vehicle Licensing Agency,
Swansea, 2006.
Gurari, N. and Okamura, A.M. Human performance in a knob-turning task, Second Joint EuroHaptics
Conference and Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems
(WHC’07), IEEE, 2007.
Hasser, C.J. and Cutkosky, M.R. System identification of the human hand grasping a haptic knob,
Proceedings of the 10th Symposium on Haptic Interfaces for Virtual Environment & Teleoperator
Systems (HAPTICS’02), IEEE, 2002.
Cándida, C. Human factors of visual and cognitive performance in driving, CRC Press, Taylor &
Francis Group, 2009.
Piechulla, W., Mayser, C., Gehrke, H. and König, W. Reducing drivers’ mental workload by means of
an adaptive man-machine interface, Transportation Research Part F 6, pp. 233-248, 2003.
Uhler, W. and Knoll, P.M. Driver Assistance Systems - driver support or driver tutelage, FISITA World
Automotive Congress 2004, Paper F2004I016, 2004.
Brännström, M., Sjöberg, J. and Coelingh, E. A situational and threat assessment algorithm for a rearend collision avoidance system, IEEE Intelligent Vehicle Symposium ‘08, Eindhoven, The Netherlands,
2008
Hollnagel, E., Nåbo, A. and Lau, I.V. A systemic model for Driver-in-Control, Proceedings of the
Second International Driving Symposium on Human Factors in Driver Assessment, Training and
Vehicle Design, pp. 86-91, 2003.
Kendel, E.S. and McRuer, D.T. A servomechanisms approach to skill development, Journal of the
Franklin Institute, pp. 24-42, 1960.
Gibson, J.J. and Crooks, L.E. A theoretical field-analysis of automobile-driving, The American Journal
of Psychology, Vol. LI, No. 3, pp. 453-471, 1938.
Westbrook, C.B. and McRuer, D.T. Handling qualities and pilot dynamics, Aero/Space Engineering,
Vol. 18, No. 5, pp. 26-32, 1959.
McRuer, D.T. and Jex, H.R. A review of quasi-linear pilot models, IEEE transactions on human factors
in electronics, Vol. HFE-8, No. 3, pp. 231-249, 1967.
McRuer, D. Human dynamics in man-machine systems, Automatica, Vol. 16, pp. 237-253, 1980.
Wier, D.H. and McRuer, D.T. A theory of driver steering control of motor vehicles, National Academy
of Sciences – National Research Council – Highway Research Record, No. 247, pp. 7-28, 1968.
McRuer, D.T. and Wier, D.H. Theory of manual vehicular control, IEEE Transactions on Man-machine
Systems, Vol. MMS-10, No. 4, 1969.
Wier, D.H. and McRuer, D.T. Dynamics of driver vehicle steering control, Automatica, Vol. 6, pp. 8798, 1970.
Wier, D.H. and McRuer, D.T. Measurement and interpretation of driver steering behaviour and
performance, SAE Preprint, International Automotive Engineering Congress, Detroit, USA, 1973.
McRuer, D.T., Wier, D.H., Jex, H.R., Magdaleno, R.E. and Allen, R.W. Measurement of driver-vehicle
multiloop response properties with a single disturbance input, IEEE Transactions on Systems, Man and
Cybernetics, Vol. SMC-5, No. 5, 1975.
McRuer, D.T., Wier, D.H. and Klein, R.H. New results in driver steering control models, Human
Factors, Vol. 19, No. 4, pp. 381-397, 1977.
Fiala, E. Lenken von Kraftfahrzeugen als kybernetische Aufgabe (in German), ATZ, 68/5, 1966.
Mitschke, M. and Niemann, K. Die Reglertätigkeit des Autofahrers bei Kursabweichungen (in German),
Deutsche Kraftfahrtforchung und Strassenverkehrstrechnik, Heft 221, 1972.
Allen, R.W., Szostak, H.T. and Rosenthal, T.J. Analysis and computer simulation of driver/vehicle
interaction, SAE technical paper series, 871086, 22 p., 1987.
Donges, E. A two-level model of driver steering behavior, Human Factors, Vol. 20, No. 6, pp. 691-707,
1978.
MacAdam, C.C. An optimal preview control for linear systems, Journal of Dynamic Systems,
Measurement, and Control, ASME, Vol. 102, No. 3, 1980.
MacAdam, C.C. Application of an optimal preview control for simulation of closed-loop automobile
driving, IEEE Transactions on Systems, Man, and Cybernetics, Vol. 11, 1981.
69
70
References
27. MacAdam, C.C. Understanding and modeling the human driver, Vehicle System Dynamics, Vol. 40,
Nos. 1-3, pp. 101-134, 2003.
28. Sharp, R.S., Casanova, D. and Symonds, P. A mathematical model for driver steer control, with design,
tuning, and performance results, Vehicle System Dynamics, Vol. 33, No. 5, pp. 289-326, 2000.
29. Sharp, R.S. and Valtetsiotis, V. Optimal preview car steering control, Vehicle System Dynamics
Supplement, Vol. 35, pp. 101-117, 2001.
30. Pick, A. and Cole, D. Neuromuscular dynamics and the vehicle steering task, Proceedings of
IAVSD’03, Kanagawa, Japan, 2003.
31. Ma, X. Driver modelling based on computational intelligence approaches - exploration and modeling of
driver-following data collected by an instrumented vehicle, PhD thesis in simulation modeling of
transport systems, KTH Architecture and the Built Environment, Stockholm, Sweden, 2006.
32. Raksincharoensak, P., Khaisongkram, W., Michitsuji, Y., Nagai, M., Shimosaka, M., Mori, T. and Sato,
T. Individual adaptation of driver assistance system based on real-world naturalistic driving database,
Part 1: Conceptual framework and driver-vehicle-environment modeling, Proceedings of AVEC ’08,
the 9th International Symposium on Advanced Vehicle Control, Kobe, Japan, 2008.
33. Raksincharoensak, P., Iijima, T., Maeda, K. and Nagai, M. Individual adaptation of driver assistance
system based on real-world naturalistic driving database, Part 2: Algorithm for detecting unusual
driving behavior based on driving database, Proceedings of AVEC ’08, the 9th International
Symposium on Advanced Vehicle Control, Kobe, Japan, 2008.
34. Fuller, R. A conceptualization of driving behaviour as threat avoidance, Ergonomics, Vol. 27, No. 11,
pp. 1139-1155, 1984.
35. Miller, B.W., Hwang, C.H., Torkkola, K. and Massey, N. An architecture for an intelligent driver
assistance system, IEEE Intelligent Vehicles Symposium Proceedings, pp. 639-644, 2003.
36. Groeger, J.A. Understanding driving: applying cognitive psychology to a complex everyday task.
Psychology Press Ltd, 2000.
37. Kiencke, U. and Nielsen, L. Automotive control systems: for engine, driveline, and vehicle, SpringerVerlag, Berlin, 2000.
38. Rumar, K. In-vehicle information systems, Int. J. of Vehicle Design, Vol. 9, Nos 4/5, pp. 548-556, 1988.
39. Regan, M.A. et al. Driver distraction – theory, effects and mitigation, CRC Press, Taylor & Francis
Group, 2009.
40. Ross, T., Midtland, K., Fuchs, M., Pauzié, A., Engert, A., Duncan, B., Vaughan, G., Vernet, M., Peters,
H., Burnett, G. and May, A. HARDIE Design Guidelines Handbook: Human Factors Guidelines for
Information Presentation by ATT Systems, Commission of the European Communities, DRIVE II
Project V2008, Workpackage No. 15, Deliverable No. L5, 1996.
41. From Brian Repa, presented by Steve Chin, GM North America, 2004
42. Wheeler, W., Lee, J., Raby, M., Kinghom, R., Bittner, A. and McCallum, M. Development of human
factors guidelines for advanced traveller information systems and commercial vehicle operations: task
analysis of ATIS/CVO functions, Research and Development, Turner-Fairbank Highway Research
Center for Office of Safety and Traffic Operations R&D, Federal Highway Administration, FHWA-RD95-176, 1996.
43. Rasmussen, J. Skills, Rules, and Knowledge; Signals, Signs, and Symbols, and Other Distinctions in
Human Performance Models, IEEE Transactions on Systems, Man and Cybernetics SMC-13, pp. 257266, 1983.
44. Ajzen, I. From intentions to actions: a theory of planned behavior, Action Control - From cognition to
behavior, pp. 11-40, 1985.
45. Goodrich, M.A. and Boer, E.R. Semiotics and mental models: modeling automobile driver behavior,
Proceedings of the 1998 IEEE ISIC/CIRA/ISAS Joint Conference, pp. 771-776, 1998.
46. Fuller, R. The task-capability interface model of the driving process, Recherche Transports Sécurité,
No. 66, pp. 47-57, 2000.
47. Jarlmark, J. Driver-vehicle interaction under influence of crosswind gusts, Licentiate Thesis,
Department of Vehicle Engineering, Royal Institute of Technology (KTH), 2002.
48. Mitschke, M. Driver-vehicle-lateral dynamics under regular driving conditions, Vehicle System
Dynamics, Vol. 22, Nos. 5-6, pp. 483-492, 1993.
49. Elander, J., West, R. and French, D. Behavioral correlates of individual differences in road traffic crash
risk: an examination of methods and findings, Psychological Bulletin, Vol. 113, No. 2, pp. 279-294,
1993.
50. Anderson, J.R. Cognitive psychology and its implications, 5th edition, Worth Publishers and W. H.
Freeman, 2000.
51. Neisser, U. Cognition and reality, W. H. Freeman and Company, San Francisco and London, 1976.
References
52. Wilde, G.J.S. The theory of risk homeostasis: implications for safety and health, Risk Analysis, Vol. 2,
No. 4, pp. 209-225, 1982.
53. Michon, J.A. Explanatory pitfalls and rule-based driver models, Accident Analysis and Prevention,
Vol. 21, No. 4, pp. 341-353, 1988.
54. Taubman-Ben-Ari, O., Mikulincer, M. and Gillath, O. The multidimensional driving style inventory –
scale construct and validation, Accident Analysis and Prevention, Vol. 36, No. 3, pp. 323-332, 2004.
55. MacAdam, C.C., Bareket, Z., Fancher, P. and Ervin, R. Using neural networks to identify driving style
and headway control behavior of drivers, Vehicle System Dynamics Supplement, Vol. 28, pp. 143-160,
1998.
56. Tricot, N., Sonnerat, D. and Popieul, J.C. Driving styles and traffic density diagnosis in simulated
driving conditions, IV'2002 - IEEE Intelligent Vehicle Symposium, Vol. 2, pp. 298- 303, 2002.
57. Barthenheier, T. and Winner, H. Das persönliche Lenkgefühl, Vortrag auf der Fahrwerk.tech 2003,
Fachtagung des TÜV-Süddeutschland, München, 2003.
58. Juhlin, M. Assessment of crosswind performance of buses, Doctoral Thesis, Department of Vehicle
Engineering, Royal Institute of Technology (KTH), 2009.
59. Nechyba, M.C. and Xu, Y. On the fidelity of human skill models, Proceedings of the 1996 IEEE
International Conference on Robotics and Automation, pp. 2688-2693, 1996.
60. Savkoor, A.R. and Ausejo, S. Analysis of driver’s steering and speed control strategies in curve
negotiation, Vehicle System Dynamics Supplement, Vol. 33, pp. 99-109, 1999.
61. Kuge, N., Yamamura, T., Shimoyama, O. and Liu, A. A driver behavior recognition method based on a
driver model framework, SAE Paper 2000-01-0349, 1998.
62. MacAdam, C.C. Development of a driver model for near/at-limit vehicle handling, Final technical
report prepared for the General Motors Corporation, 2001.
63. Underwood, G., Chapman, P., Berger, Z. and Crundall, D. Driving experience, attentional focusing, and
the recall of recently inspected events, Transport Research Part F, 6, pp. 283-304, 2003.
64. Wier, D.H. and Allen, R.W. Psychological and response measurements in driving tasks, SAE Paper No.
720139, 1972.
65. Nordmark, S., Jansson, H., Palmkvist, G. and Sehammar, H. The new VTI driving simulator. Multi
purpose moving base with high performance linear motion, Proceedings of DSC Europe ‘04, Paris,
France, 2004.
66. VTI. World-class simulator, Official folder for VTI Simulator III, http://www.vti.se/5053.epibrw,
[accessed 8/12/2008].
67. VTI. Simulator III, Official Web-presentation of VTI Simulator III, http://www.vti.se/templates/...
Page____11861.aspx, [accessed 8/12/2008].
68. Mechanical Simulation Corporation, http://www.carsim.com, [accessed 28/3/2010].
69. Mechanical Simulation Corporation. Math model features, CarSim Manual, Version: 6.05, Release:
January 2006.
70. Pacejka, H.B. and Sharp R.S. Shear force development by pneumatic tyres in steady state conditions. A
Review of Modeling Aspects. Vehicle System Dynamics, Vol. 20, Nos. 3-4, pp. 121-176, 1991.
71. Abe, M. Vehicle handling dynamics – theory and application, Butterworth-Heinemann, Elsevier Ltd.,
2009.
72. Pacejka, H.B. Tyre and vehicle dynamics, Oxford : Butterworth-Heinemann, 2002.
73. Gillespie, T.D. Fundamentals of vehicle dynamics, Society of Automotive Engineers, Inc., 1992.
74. Gou, K. and Guan, H. Modelling of driver/vehicle directional control system, Vehicle System
Dynamics, Vol. 22, Nos. 3-4, pp. 141-184, 1993.
75. Iguchi, M. A study of manual control, Journal of Mechanic Society of Japan, Vol. 62, No. 481, 1959 (in
Japanese) [72].
76. Iguchi, M. et al. A fundamental study of driver behaviour for automobile simulation, Symposium of
SAE-Japan No. 6, 1960 (in Japanese) [74].
77. McRuer, D.T. and Krendel, E.S. The human operator as a servo system element, Journal of the Franklin
Institute, Vol. 267, No. 5, pp. 381-403, No. 6, pp. 511-536, 1959.
78. Kondo, M. and Ajimine, A. Driver’s sight point and dynamics of the driver/vehicle system related to it,
SAE paper No. 680104, 1968.
79. Chatelet, P. and Anhtuan, P. Dynamique du système conducteur-véhicule, Ingenieurs de l’automobile,
pp. 757-812, 1973 (in French) [82].
80. Legouis, M., Laneville, P., Bourassa, P. and Payre, G. Characterization of dynamic vehicle stability
using two models of the human pilot behaviour, Vehicle System Dynamics, Vol. 15, No. 1, pp. 1-18,
1986.
81. Legouis, M., Laneville, P., Bourassa, P. and Payre, G. Vehicle/pilot system analysis: a new approach
using optimal control with delay, Vehicle System Dynamics, Vol. 16, Nos. 5-6, pp. 279-295, 1987.
71
72
References
82. Plöchl, M. and Edelmann, J. Driver models in automobile dynamics application, Vehicle System
Dynamics, Vol. 45, Nos. 7-8, pp. 699-741, 2007.
83. Yoshimoto, K. Simulation of driver/vehicle system including preview control, Journal of Mechanics
Society of Japan, Vol. 71, 1968 (in Japanese) [74].
84. Yoshimoto, K. Simulation of driver/vehicle system, Automobile Technique, Vol. 25, No. 10, 1971 (in
Japanese) [74].
85. Yoshimoto, K. A self paced preview tracking control model of an automobile driver, Control Science
and Technology for the Progress of Society, Proceedings of the 8:th Triennial World Congress, Vol. 15,
1981.
86. Sheridan, T.B. Three models of preview control, IEEE transactions on human factors in electronics, Vol.
HFE-7, No. 2, 1966.
87. Berardinis, L.A. What you need to know about fuzzy logic and neural networks, ISA, pp. 1131-1140,
1993.
88. Neusser, S., Nijhuis, J. and Spaanenburg, L. Developments in autonomous vehicle navigation,
CompEuro'92 Computer Systems and Software Engineering Proceedings, pp. 453-458, 1992.
89. Ohno, H. Analysis and modeling of human driving behaviors using adaptive cruise control, Industrial
Electronics Society, 2000. IECON 2000. 26th Annual Conference of the IEEE, pp. 2803-2808, 2000.
90. Matlab script by Sigworth, F. State-machine simplex minimizer, Matlab central file exchange,
http://www.mathworks.com/matlabcentral/fileexchange/, [accessed 15/7/2008].
91. Caceli, M.S. and Cacheris, W.P. Fitting curves to data – the Simplex algorithm is the answer, Byte, Vol.
9, No. 5, pp. 340-362, 1984.
92. Helmers, G. and Törnros, J. Effect of transition curves on driver’s safety margins and the influence of
training – an experiment in a driving simulator, VTI, R501, 2006 (in Swedish).
93. Helmers, G. and Törnros, J. Effect of vehicle speed and side acceleration related to the location of the
centre of rotation for banking of sharp curves, and influence of training – an experiment in a driving
simulator, VTI, M967, 2004 (in Swedish).
Nomenclature
Notation
ay
Lateral acceleration [m/s2]
aR
Radial (centripetal) acceleration [m/s2]
b
Distance from the vehicle’s centre of gravity to the rear axle [m]
B(s)
Driver feedback (i.e. prediction) function
C
Set of ordered consequences
CB
Curve beginning
C12 , C 34
Cornering stiffness coefficient, front and rear axle [N/rad]
D
Derivative operator
f
Distance from the vehicle’s centre of gravity to the front axle [m]
F12 , F34
Lateral force, front and rear wheels respectively [N]
G ( s)
Transfer function for the vehicle system
H (s )
Transfer function for the driver compensation
HVS
Number of high (or low) skill test subjects in high value section
Jz
Inertia [Nm·s2]
K
Gain [1]
K us
Understeer gradient [rad·s2/m]
Kd , K p , Ki
Driver model coefficients (derivative, proportional, and integration)
L
Wheel base [m]
LVS
Number of high (or low) skill test subjects in low value section
m
Mass [kg]
M SW
Steering wheel torque [Nm]
P( s)
Driver preview strategy
R
Curve radius [m]
73
Nomenclature
74
R min
Smallest curve radius [m]
s
Laplace derivative operator
SS
Size of sections in grade calculation
t
Time [s]
td
Brain response delay [s]
th
Driver action delay [s]
TL
Lead [s]
Tl
Lag [s]
Tp
Preview time [s]
U
Set of decisions or actions
V
Velocity [m/s]
vx , v y
Longitudinal and lateral velocity [m/s]
v& x , v& y
Time derivative of longitudinal and lateral velocity [m/s2]
vT , v R
Tangential and radial velocity [m/s]
y road
Lateral position, road reference [m]
yp
Predicted lateral position [m]
y (t )
Feedback vector
β
Body slip angle [rad]
δ
Steering angle [rad]
δA
Ackermann steering angle [rad]
δ SW
Steering angle [rad]
θ
Perceived state of the environment
µ
Friction coefficient [1]
ψ
Yaw (heading) angle [rad]
ψ&
Yaw rate [rad/s]
ψ&&
Yaw acceleration [rad/s2]
ψ road
Yaw angle, road reference [rad]
Nomenclature
ωc
75
Cross-over frequency [Hz]
Abbreviations
3D
3-dimensional
ACC
Adaptive cruise control
BA
Behaviour actuation
ba
Braking to avoid collision
CG
Centre of gravity
CH
Correct and hold
CO
Continuously oscillating
DA
Discriminant analysis
DiC
Driver in control
DLC
Double lane change
DOF
Degrees of freedom
ECG
Electrocardiogram
GM
General Motors
HC
Heading control system
HMM
Hidden Markow model
ISO
The International Organization for Standardization
JDVS
Joint driver-vehicle system
KB
Knowledge-based
KTH
The Royal Institute of Technology
MBS
Multi-body system
MC
Mixed compensating
M.C.A
Multiple correspondence analysis
MM
Mental model
OC
Over-compensating
ODE
Ordinary differential equations
QC
Quickly compensating
RB
Rule-based
RMS
Root-mean-square
Nomenclature
76
SB
Skill-based
SC
Slowly compensating
SP
Sensor perception
sr
Speed regulation
tr
Time headway regulation
UC
Under-compensating
UMTRI
The University of Michigan Transport Research Institute
VINNOVA
The Swedish Agency for Innovation Systems
VTI
The Swedish National Road and Transport Research Institute
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

Related manuals

Download PDF

advertisement