A Review of the Techniques for Indoor Location based Service

A Review of the Techniques for Indoor Location based Service
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
A Review of the Techniques for Indoor Location based Service
Gyeyoung Lee and Jaegeol Yim
Dongguk University at Gyeongju, Korea
{lky, yim}@dongguk.ac.kr
Abstract
A location based service (LBS) provides useful information to the users based on the
geographic location the user designates or the user is currently located. Examples of LBS
include directory service, gateway service, location utility service, presentation service, route
service, and so on. These services are very useful to the users, and LBS should be available
indoor: subways, large shopping malls, department stores, factories, and so on. This
manuscript summarizes the techniques needed in development of indoor LBS (ILBS) including
indoor positioning, indoor moving objects database system, rendering drawings, and web
service.
Keywords: Indoor Location Based Service, Indoor Positioning, Fingerprints, K-NN, Web
Services, Kalman filter
1. Introduction
A location based service (LBS) is to provide useful information to the users based on
the geographic location the user designates or the user is currently located. Examples of
LBS include directory service, gateway service, location utility service, presentation
service, route service, and so on [1]. These services are very useful to the users and
should be provided not only outdoor but also indoor. Those LBSs provided indoor is
called ILBS (Indoor Location Based Service). However, most of the existing LBSs are
outdoor ones and they use GPS to determine user‟s current location. In the case of
indoor, GPS signal cannot be utilized and there is no economic way of determining
user‟s location even though there are many research results about indoor positioning
have been published.
This manuscript surveys the fundamental techniques for ILBS development inc luding
positioning, database management, rendering drawings, and web services. LBS cannot
be realized without the information of user‟s location positioning, determining user‟s
location, is the most important and essential ingredient of LBS. Database management
is another important ingredient because there are many moving, pedestrians for
example, and stationary, stores for example, objects involved in an ILBS system.
Rendering drawings is another important research subject in the field of ILBS because a
drawing is the most important component of the user interface of an ILBS system. For
LBS, the map is used as the main component of the user interface, but in the case of
ILBS, a drawing is an unpublished private property. The solutions for those problems of
positioning, database management and rendering drawing should be published as web
services so that other application developers can utilize them in their development .
1
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
2. Related Works
There are many positioning systems including GPS [2], wide-area cellular-based
systems [3], infrared-based system [4], radio frequency (RF) + ultrasonic-based systems
[5, 6], physical contact systems [7], various computer vision systems [8], and RF based
systems [9, 10]. Among them RF-based wireless LAN positioning techniques are most
interesting in the field of indoor positioning because of the following reasons. GPS
signal is not available inside of a building so GPS system cannot be an indoor
positioning system and the others require special equipments dedicated for p ositioning.
Whereas RF-based wireless LAN positioning systems do not require additional
hardware dedicated for positioning and wireless network is being serviced everywhere
including college campuses, airports, hotels and homes [11].
A moving object database (MODB) is a database representing information on moving
objects, particularly their locations. The purpose of the MODB is to provide answers to
various spatiotemporal information queries such as the following: “Where am I?” “Who
is the closest friend to me?” “How far is he?” “When and where did I meet Kim?” A
location-based service (LBS), such as the fleet management system for a taxi company
or a logistics company, cannot be realized without the availability of spatiotemporal
information. Therefore, MODB techniques have been developed in the LBS field.
For most MODB applications, the moving object is equipped with a GPS (Global
Positioning System) unit that sends information on the object‟s location, time, and
velocity, among others, to the MODB. However, GPS signals are usually unavailable
inside enclosed structures; thus, MODB techniques used for outdoor LBSs cannot be
directly applied to indoor LBS systems. An indoor moving object can be a mininotebook computer, a PDA, or a smart phone carried by a pedestrian. The physical care
covered by indoor moving objects is much smaller than that by outdoor ones, and the
speed of indoor moving objects is much slower than that of outdoor ones.
Indoor LBSs are much more useful than outdoor LBSs. Indoor LBSs may include an
electronic museum guide, a shopping concierge at a department store, asset tracking,
VoIP-911 caller location identification, and so forth. An electronic museum guide
pushes the digital content related to the exhibit that the user is viewing, sugge sts points
of interest, and provides the information the user requests [12]. The following summary
is also from [12].
In the field of MODB, models of location information, techniques of uncertainty
management, query languages accessing spatiotemporal data, indexing techniques, data
mining, and security have to be studied. Wolfson has reviewed these topics in [13]. The
location information in a MODB is naturally uncertain. Olston introduced a method of
handling the uncertainty [14] and Tao et al. introduced “probably restricted rectangle”
and the information retrieval method using the rectangle [15] .
The positions of the spacecrafts are important information in national defense
system. [16] introduces their design of moving object database system for military
command. The main elements of the system include a traditional relational DBMS,
spatial wrapper of the DBMS, API component between GUI and the DBMS. [17] is
another MODB prototype implemented on top of a traditional relational DBMS. [18]
proposes a CASE tool supporting spatiotemporal conceptual modeling for moving
object database applications. It has the following two advantages: (1) It supports both
spatiotemporal and non-spatiotemporal characteristics and compatible with the ER
model. (2) It supports systematic spatiotemporal characteristics and can represent
2
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
different spatiotemporal changes for diverse moving objects applications. [19] explains
how to use SECONDO in building a moving objects database system.
The query based indexing techniques are about indexing queries of the moving
objects and not about indexing the location of these objects. The location based
indexing techniques can be divided into two categories, indexing the current and future
positions and indexing the past positions. These indexing methods are all variants of RTree. [20] introduces Delineated R-Tree (DR-Tree) indexing structure which is highly
balanced and has performance advantages over other R-tree based indexing methods.
[21] proposes another new indexing method, S-TB Tree. [22] proposes the indexing
method of efficiently processing historical spatiotemporal pattern queries ,
There is no commercial moving object database in the markets yet. Applications of
MODB inevitably provide a map on their user interface. [23] introduces a digital map to
indicate locations of interested static or moving objects. They use Oracle Spatial
SDOAPI Java Class Library to handle spatial data. [24] presents a formal model where
the geometric components of the thematic layers in a GIS are represented a s an
OLAP(On Line Analytical Processing) dimension hierarchy and introduces the notion
of spatial aggregation. Then, they also address moving object aggregation over a GIS .
The advances in Global Positioning Systems, wireless communication systems and
miniaturization of computing devices have brought an emergence of various
applications in Location-Based Services (LBS). As a result, there is an increasing need
for efficient management of vast amounts of location-in-time information for moving
objects. Therefore, one of the most important issues a MODB must address is spatio temporal languages. The queries MODB should be able to handle include finding the
similar trajectories of a moving object. With this kind of query, we can determine
migration patterns of animals and recommend a route to a driver. Nearest Neighbor
(NN) query is defined as follows: Given a set of trajectories T and a trajectory Q, find
the trajectory in T which has the smallest distance to Q. If we find 2 closest trajectories
instead of just 1 then the query is called 2NN. [25] addresses Set Nearest Neighbor
query. An example KNN(K-nearest neighbor) query is “What will be the query‟s 3NN 2
minutes from now?” [26] introduces a new multi-threaded and cache-conscious
algorithm to efficiently process massive continuous KNN queries. [27] proposes
techniques including pruning candidate objects for processing a nearest neighbor query
when the location of the query object is specified by an imprecise Gaussian distribution.
[28] proposes a Frechet distance based approach to similarity join for large sets of
moving object trajectories and shows that the proposed method is 50% faster than the
traditional methods. There are two kinds of moving objects: moving points like vehicles
and moving regions like hurricanes and oil spills. Spatial objects evolve as time flows
and their topological relationships develop over time. Spatio-temporal predicates have
been proposed to ask for these time-varying relationships. [29] proposes a generic
algorithmic scheme that can evaluate spatio-temporal predicates. The following is an
example continuous spatio-temporal query: “Continuously, inform commander C with
all friendly units that are within ten miles from soldier S.” To minimize the execution
cost of this kind of queries, [30] proposes Query-Track-Participate(QTP) query
processing model where a query is continuously answered by a querying server, a
tracking server, and a set of participating servers. [31] introduces an algebraic Spatio Temporal Trajectory data type (STT) for the representation of object trajectory. The
STT is an abstract data type endowed with a set of operations designed as a way to
cover the syntax and semantics of a given trajectory.
3
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
Updating policies for MODB are thoroughly discussed in [32]. They introduced three
moving object database updating policies, Speed Dead-Reckoning (sdr), adaptive dead
reckoning (adr), and disconnection detecting dead-reckoning (dtdr). A dead-reckoning
update policy updates the moving object‟s database location only when th e difference
between its actual location and its database location is greater than the threshold th. The
threshold th is a constant during the trip in the case of sdr, whereas it is newly
computed whenever the database location is updated in the case of adr. In the case of
dtdr, the threshold th is not only newly computed whenever the database location is
updated but it also decreases as the time interval since the last update increases .
Typical applications of MODB include intelligent transport systems, di gital
battlefield, location e-commerce, and so forth. A query that must be evaluated for a
time interval, “List the nearest gas stations to my vehicle in the next 15 minutes” for
example, is called a continuous query. Existing MODBs work well for instantan eous
queries but not for continuous ones. [33] proposes a new updating strategy which
improves the quality of continuous query results. Then, it provides their simulation
results to show the effectiveness of their updating strategy. In their simulation, th ey use
a standardized random number generator to assign velocity to the moving object. This
implies that they assume that a moving object is equipped with a device like GPS which
provides the velocity of the moving object. In the case of indoor positioning there is no
device which provides the velocity of the moving object.
The estimated values are used to predict the moving object‟s future position. For the
future prediction, most techniques use some mathematical formulas of motion derived
from its recent movements. However, an object‟s movements are not that simple.
Therefore, an object‟s trajectory patterns are used for prediction recently. [34]
introduces a hybrid prediction model where both mathematical formulas and trajectory
patterns are considered. [35] proposes an asynchronous position updating algorithm for
continuous queries of moving objects which improves accuracy, reduces the
communication cost, and balances server load more evenly. An example of continuous
query is “All vehicles which will appear in the region R in the next 10 minutes.”
Processing the location stream which is faster than the database can handle is one of
the problems we have to solve in order to build a high performance MODB. [36]
proposes a location filter in order to address the problem of processing the location
stream. The location filter consists of the on-line location filter and the temporal
location manager. The on-line location filter executes the filtering algorithm plug-ins
and filters the location stream on-line while the temporal location manager insert
location data into the database at an acceptable insertion rate.
Techniques of rendering and manipulating a map have been developed in the field of
GIS. Nowadays, many organizations provide open APIs for map manipulati on.
However, these are all for outdoor LBS development [37, 38]. This manuscript
introduces the module of rendering drawings we have implemented.
Web services are typically application programming interfaces or Web APIs that are
accessed via Hypertext Transfer Protocol and executed on a remote system hosting the
requested services [39]. This manuscript introduces some web services providing
essential functions for ILBS development.
3. Our Research Results
In this section, the techniques we have developed will be introduced. The materials
in this section are from the papers we have published.
4
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
3.1. Indoor Positioning
3.1.1. K-Nearest Neighbor Method: In K-NN, we build a look-up table in the first phase,
or off-line phase. The entire area is covered by a rectangular grid of points called
candidate points. At each of candidate points we measure the RSSs many times. Let
RSS ij denote the j-th received signal strength of the signal sent by AP i. A row of the
look-up table is an ordered pair of (coordinate, a list of RSSs). A coordinate is an
ordered pair of integers (x, y) representing the coordinates of a candidate point. A list
of signal strengths consists of five integers, RSS 1 , RSS2 ..., where RSS i is an average of
signal strengths RSS ij received at (x, y) and sent by AP i. An example of look-up table is
shown in Table 1.
In the second phase, or on-line phase, the positioning program gathers RSSs the user
receives at the moment. If the positioning program is running on the user‟s handheld
terminal, then the terminal itself will collect RSSs. Let X=(-40, -56, -54, -69, -66) be
the vector of the collected RSSs. K-NN, then look up the look-up table and finds the
closest candidate point, CP2 in the case of Table 1, and returns it as the user‟s current
location. If K equals 2, then it will find two closest candidate points and return the
average of them as the user‟s current location.
Table 1. An Example Look-up Table of K-NN (C.P stands for Candidate Points,
CPi is coordinates of i-th C.P, APi is the MAC address of i-th AP)
AP
C.P
CP1
CP2
CP3
...
AP1
AP2
AP3
AP4
AP5
-39
-40
-44
...
-55
-56
-42
...
-56
-55
-62
...
-70
-69
-45
...
-67
-66
-61
...
3.1.2. Decision Tree for Positioning: In the off-line phase of decision tree method, we
build a decision tree with training data. An example training data set is shown in Table
2. Table 2 is similar to Table 1. Only differences are 1) RSSs are not averages, 2) RSSs
are discretized. An example of discretizing policy can be I1 = {x| x>-30}, I2 =
{x| - 40  x  -30 }, I3 = {x| - 50  x  -40 }, ....
Given a set of training data, we build a decision tree with the algorithm
Construct_DT shown in Fig. 1. Step (4) of the algorithm computes I where I is the
expected information needed to classify a given sample and is given by
m
I ( s1 , s2 ,..., sm )   pi log 2 ( pi )
i 1
where, m is the number of candidate points, S is the number of tuples in the training data set
(rows of Table in the algorithm), si is the number of rows of S in class CPi , pi 
si
s .
Step (5) of the algorithm computes the entropy, or expected information based on the
partitioning into subsets by CPi . Let CPi have v distinct values, {a1 , a2 , ..., av } . CPi can
be used to partition S into v subsets, {S1 , S 2 , ..., Sv } , where S j contains those samples in
5
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
S that have value a j of CPi . Let S ij be the number of samples of class CPi in a subset
S j . The entropy E( CPi ) is given by
v
sij  ...  smj
j 1
s
E (CPi )  
I ( sij ,..., smj ).
At step (6) the algorithm computes information gain G( CPi ) by the following expression,
Gain(CPi )  I (s1 , s2 , ..., sm )  E (CPi ) .
Table 2. An Example Training Data Tuples (C.P stands for Candidate Points,
CPi is coordinates of i-th C.P, APi is the MAC address of i-th AP, I stands for
interval)
AP
C.P
CP1
CP2
AP1
AP2
AP3
AP4
AP5
I2
I1
...
I3
I1
I1
I2
I2
I5
I1
I5
I1
I2
I3
I1
I2
...
...
...
An example decision tree is shown in Figure 1. Depth of a decision tree is the
number of APs, or N, and the number of branches of a node is the number of intervals,
or I. A leaf of a decision tree is labeled with CPi , or candidate point. Given a vector of
measured RSSs, X=(RSS1, RSS2, ...), the on-line phase of decision tree method, when
the decision tree built in the off-line phase looks like the one in Fig. 1, examines RSS2
because the root of the tree is labeled AP2. If RSS2 belongs to I2, then it will take the
second branch and will examine RSS4 since the second child node of the root is labeled
AP4. For each node, there are at most I branches and the depth of a decision tree is N.
Therefore, the time complexity of the on-line phase of decision tree method is O(I*N).
Figure 1. An Example Decision Tree where I=6, N=5, and M=96.
6
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
Experimental Evaluation
For experimental evaluation of accuracy of the methods, we have implemented KNN, Bayesian and decision tree methods on a SAMSUNG SENS R50 laptop computer
equipped with Intel(R) PRO/Wireless 2200BG Network Connection as its WLAN card
using Microsoft Visual C# 2005 Express Edition Beta. It is known that neural network
gets worse accuracy than K-NN and we did not implement neural network method. The
experimental results are presented in this section.
Test Bed
The test bed for the experiments is the Micro LAB on the 4-th floor of Natural
Science Building as shown in Figure 2.
Figure 2. The Test Bed
Test Results
We first performed experiments to see if the number of samples for one entry (the
average of the samples) of Lookup table is related to the accuracy of K -NN. With the
results shown in Figure 3, we decided to use the average of 30 samples as an entry of
Lookup table. We have used X to denote the vector of RSSs measured in the on -line
phase. We obtained X with the average of 10 measures for this experiment.
Figure 3. The Number of Samples for an Entry of Lookup Table and Accuracy
7
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
We have used X to denote the vector of RSSs measured in the on -line phase. We
performed experiments to see how the number of samples for X is related to accuracy.
Figure 4 shows the result of the experiments. An entry of the lookup table for this
experiment was the average of 100 samples. With the result we can conclude that if we
use the average of 10 samples as X then the accuracy will be much improved.
Figure 4. The Number of Samples for X and Accuracy
We performed experiments of running 1-NN, Bayesian and decision tree methods on
training data of N=5, I=6, and M=96 in order to compare their accuracies. The test
results are shown in Figure 5. In the figure, „number of samples‟ is the same as the
number of measurements we have performed for each entry of Lookup table. An entry
of Lookup table is the average of the measurements. When the number of samples is 10,
K-NN method is much better than others. However, the difference gets decreases as the
number of samples increases and when the number of samples is 50 the accuracies of
the three methods are almost same. For this experiment, X was the average of 10
measurements.
Figure 5. Accuracy of Three Methods
3.1.3. Trilateration: If we measure N ranges,
n1  ( X1 Y1 Z1 ) ,
T
8
, nN  ( X N YN Z N )
r1 , r2 , ..., rN from N base stations,
T
to a mobile terminal,
m  ( x y z )T as shown in Figure 6,
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
2
r
then we can estimate the coordinates of m by using trilateration. By squaring, i can be expressed as
follows:
( x  X i )2  ( y  Yi )2  ( z  Zi )2  ri2 , ( for i  1, 2,..., N )
r2
n2
r1
n1
m
n3
r3
Figure 6. A Diagram to Illustrate Trilateration
By subtracting r1 from ri (i  2,
2
2


, N ) , we have Ax  b , where
 x
 ( X 2  X 1 ) (Y2  Y1 ) ( Z 2  Z1 ) 

 , x   y
A  2
 


 z 
( X N  X 1 ) (YN  Y1 ) ( Z N  Z1 ) 
 ( X 22  X 12 )  (Y22  Y12 )  ( Z 22  Z12 )  ( r22  r12 ) 


b

( X N2  X 12 )  (YN2  Y12 )  ( Z N2  Z12 )  ( rN2  r12 ) 


.
When the coordinates are 3 dimensional, we need to have at least 4 base stations. By
applying the MMSE (Minimum Mean Square Error) method, we can obtain x̂

(approximation of x ) with the following position estimates:
xˆ  ( AT A)1 AT b --------------- (Expression 1)
3.1.4. Extended Kalman Filter: The Kalman filter iteratively predicts the position of the
mobile terminal and updates the prediction with new measurements to yield an estimate.
In positioning, the measurement equation is represented as a nonlinear model, and
linearization should be performed to derive a linear equation. The extended Kalman
filter (EKF) considers the real-time linearization of the system function at the previous
state estimate and that of the observation function at the corresponding predicted
position. The measured distances, ri , can be expressed as follows:
ri  ( X i  x)2  (Yi  y )2  ( Zi  z )2  vi
9
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
where vi represents the measurement noise and is assumed to be white Gaussian
noise (AWGN) with a normal probability distribution of 0 mean and  i2 variance.
Using the Taylor approximation at a nominal point x0  ( x0 y0 z0 )T , we have:
ri  ri 0 
ri
x
 x  vi  hxi  x  hyi  y  hzi z  vi
x  x0
where ri 0  ( X i  x0 )2  (Yi  y0 )2  ( Zi  z0 )2 is the computed distance between the
nominal point and the i-th base station, (hxi 
x0  X i i y0  Yi i z0  Z i
, hy 
, hz 
) is the
ri 0
ri 0
ri 0
LOS (Line Of Sight) vector from the base location to the i-th base station, and
 x  ( x  y  z ) is the position error vector to be determined. For N base stations, we
have the following expression:
 r1  r10   h1x

 


 rN  rN 0   hxN

 
h1y
hyN
hz1   x   v1 
   
  y    
N
hz    z  vN 
This equation can be rewritten in the following form, where subscript k is the index
for the discrete time sequence:
 rk  rk  r0  H k xk  vk
----------- (Expression 2)
The measurement noise is AWGN with vk ~ N (0, Rk ) . Here, Rk  diag ( i2 ) where
diag stands for diagonal matrix. If we have more than 3 measured distances, we can
estimate the position of the mobile terminal by using the WLSE (Weighted Least
Squares Estimate). However, by adding system models to the above measurement
models and applying the Kalman filtering process, a more reliable position can be
found. The P (Position), PV (Position Velocity) and PVA (Position Velocity
Acceleration) models are generally used in navigation as a system model. For
positioning, we can use the P model described by the following expression:
xk 1   k xk  wk
where  k is the state transition matrix and wk ~ N (0, Qk ) the modeling error, or the
system error. Table 3 summarizes the EKF processing. The estimate we can obtain with these
expressions is the value of  xˆk , and the final solution, i.e. the user‟s location, can be
obtained with the following expression:
xˆk  x0   xˆk .
10
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
Table 3. A Summary of the EKF Processing
Linearized state model: xk 1   k xk  wk , wk ~ N (0, Qk )
Linearized measurement model: rk  r0  H k xk  vk , vk ~ N (0, Rk )
1) Initial guess: x0  E ( x0 ) and P0  var( x0 )
2) Linearizing:
 ( X  x  )2  (Y  y  ) 2  ( Z  z  ) 2 
1
k
1
k
1
k


rk  r0  H k xk  vk , r0  



 2
 2
 2
 ( X N  xk )  (YN  yk )  ( Z N  zk ) 
3) Kalman Gain: Kk  Pk H kT ( H k Pk H kT  Rk )1
4) Measurement update: xˆk  xˆk  Kk ( rk  r0 )
5) Update error covariance: Pk  ( I  K k H k ) Pk
6) State propagation: xˆk1   k xˆk , Pk1   k Pk Tk  Qk
7) Go to step 2
Positioning is a special case of the EKF process where  k  I and Qk  0 . It replaces
step 6 in Table 3 by the following simplified expression:
xˆk1  xˆk , Pk1  Pk
The implementation of a positioning system using trilateration or the extended
Kalman filter requires the relationship between the distances and RSSIs. For this
purpose, we read the RSSI every 1m from an AP 300 times and, based on this
information, plotted the relationship between the distance and RSSI, as shown in Figure
7. Using this relationship, we can obtain the distance to the AP based on the RSSI from
the AP.
Relation of distance and RSSI
-20
Mean of measurements
Propagation model
RSSI (dBm)
-30
-40
-50
y = -12.1ln(x) - 33.66
-60
-70
-80
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
Distance (m)
Figure 7. Relation between Distance and RSSI
11
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
Experimental Analysis
The graphical user interface (GUI) of the trilateration-based positioning system is
shown in Figure 8. The floor plan shown in Figure 8 represents the 4 th floor of the
Natural Science Building, the test bed. When the button, “Location?”, is clicked, the
system runs the trilateration algorithm and writes the X and Y coordinates returned
from the algorithm in the box labeled “Position Info :” It also draws a big dot at the
position (X, Y) on the floor plan. When the button, “Test”, is clicked, it provides dialog
boxes in which the actual user‟s location and the integer N can be input. It then runs the
trilateration algorithm N times and records the results in the file named Result. We also
implemented the EKF positioning method, and its GUI is similar to that of trilateration based positioning system shown in Figure 8.
Figure 8. A Typical GUI of the Positioning Systems
There are many parameters affecting the efficiency of a positioning system and the
number of APs with significant signal strengths is one of them. If the RSSI fro m an AP
is less than -70 then the distance to the AP cannot reliably be determined and we
consider it to be insignificant. In the following experiments, we had 4 APs with
significant signal strengths.
They also performed experiments using the trilateration and EKF positioning algorithms at
40 different locations. At each location, we repeated the iteration of estimating the position
300 times. The results of the experiments are summarized in Figure 9. The X-axis represents
the number of iterations and the Y-axis represents the errors of the positioning results in
meters. A „+‟ in the graph represents an error of trilateration positioning and is the average of
the errors obtained at the 40 different locations. The dashed line (the points are so close that it
appears to be a solid line) represents the errors of the EKF positioning and the straight line
represents the average error of the trilateration (ET: from now on ET stands for “average error
of the trilateration”) positioning. As the graph shows, the average of the errors of the
trilateration is 4.07 m, whereas the error of EKF (EE: from now on EE stands for “average
error of the EKF”) converges to 3.528m. There is a difference between ET and EE. This
12
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
difference comes from the fact that ET only counts distances which are always positive. For
example, consider the case where the estimates are (140, 200) and (160, 200). The
trilateration would determine the user‟s position to be (150, 200) and ET would be 10. On the
other hand, the estimate of EKF would be close to (150, 200) and EE would be close to 0.
Figure 9. Comparison of Trilateration vs. EKF
3.2. Moving Object Database for ILBS
The major components of an indoor MODB system are the database server, the web
server and the communication server. The communication server continuously receives
(Time, Location) tuples from the mobile terminals. The Location is the coordinates of
the measured location of the mobile terminal at the moment of Time. The
communication server passes the received tuples to the database server. The web server
delivers the user‟s commands to the database server and displays the answers from the
database on the user‟s terminal. The database server stores all the (Time, Location)
tuples from all the mobile terminals in the application field of the MODB. The database
server also manipulates the queries from the web server and returns the answers of the
queries to the web server referring to the digital map of the domain of the application .
[12] proposes an updating strategy with which we can save the communication cost
without degrading the accuracy of database locations. The strategy applies the Kalman
filter on the positions measured at t0 , t1 , …, t i in order to estimate the state of the
moving object and extrapolates the positions at the moments of ti+1, ti+2, …, with the
estimated state. If the difference between the extrapolated position and the measured
position is not greater than the given threshold then the strategy omits sending the
measured position. The strategy performs the extrapolation when the elements of P are
all less than 0.001.
The algorithm of the strategy is shown in Table 4. For the first step, the algorithm applies
the Kalman filter process on the consecutive i recently measured positions. We are using the
decision tree method for the measurement. The formats of the matrices used in the Kalman
filter process are as follows.
 xk 
1
y 
0
k ,

Xk 

vxk  A  0
 

0
v yk 
0 t
1 0
0 1
0 0
0
1
t  , H  
0
0

1
0 0 0
1 0 0
13
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
The initial values for the other matrices are determined as follows.
0
0
0 
0.001
1
,
 0
0
.
001
0
0  R  
Q
0
 0
0
0.001
0 


0
0
0.001
 0
0 ,
1
0
0 
300 0
 0 300 0
0  .
P0  
 0
0 300 0 


0
0 300
 0
After the first step, the algorithm checks the elements of P , and if they are all less
than a small constant „Epsilon‟ (0.001 for example) then the algorithm performs the
extrapolation. In the process of the extrapolation, we replace all the t s of A with 1s.
Then, we multiply the A by the i th state obtained by the first step. The result of the
multiplication is assigned to the variable „prediction‟ in the algorithm. At this moment,
if the difference between the i+1 th measured location and the prediction is less than the
threshold (3 meters, for example) then we skip sending the measured location to the
communication server. Otherwise, we do not skip sending the measurement.
Table 4. The Algorithm for Updating
1) Apply the Kalman filter process shown in Figure 1 on the i measured locations to
produce x̂i
2) measurement = i+1th measurement obtained by the positioning algorithm;
3) if (elements of P < Epsilon) then {
4)
prediction = Axˆi ;
5)
If (|prediction – measurement| < Threshold)
6)
Then skip sending
7)
Else send measurement to the server
8) Else send measurement to the server: }
If we skipped sending the i+1 th measured location then we keep applying the
updating algorithm on the i+2 th measurement, i+3 th measurement, and so on. When we
apply the algorithm on the i+2 th measurement, the t s of A are all set to 2. When we
apply the algorithm on the i+3 th measurement, the t s of A are all set to 3, and so on.
However, the value of i does not change. In other words, we always apply the Kalman
filter on the i recent measurements.
Sometime later, it happens that (|prediction – measurement| < Threshold) is no longer
true. If (|prediction – measurement| < Threshold) is not true for n (n=3, for example)
consecutive measurements, then we conclude that our prediction is no longer valid and
let the mobile terminal send the n measurements to the server.
When the database server receives a query of “Where was the person A at the time
xx?” it looks for (xx, Location) on the database. If it finds one then it returns the
Location, otherwise it performs our Kalman filter based extrapolation and returns the
result as the answer for the query. Since our updating policy skips updating when the
difference between the prediction and the measurement is less than the Threshold, the
answers returned by the database server are always within the Threshold from the
measured location.
14
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
Our experimental results are summarized in Table 5. When we set Threshold to 3
meter, our strategy skipped sending 74.6 times in average. That is 52.5% of the
measurements. Referring to Table 5, we can notice that the average error of our
measured positions is 2.924. That is very close to our Threshold. That explains why the
skip rate of our experiment is 52.5%. We can also notice that the average error
(3.05214) of recorded positions for the experiment is a little higher than that (2.924) of
“without updating policy” but they are very close. This is natural because we start
submitting measurements when “(|prediction – measurement| < Threshold)” holds.
In the other set of experiments, we set Threshold to 2 meter. Our strategy only
skipped sending 37 times, or 26%, in average. It is natural that the smaller value for the
Threshold results in the smaller number of skips. As the number of skips gets smaller,
the average of the errors of the recorded measurements for this experiment gets closer
to that of “Without Updating Policy.”
With these experimental results, we can conclude that our updating method c an save
about a half of the communication cost without almost any tradeoffs for indoor moving
objects databases when the moving objects measure their positions with the WLAN
based K-NN positioning method.
Table 5. Summary of Experiments
Number of skips
Average of errors (meter)
Without Updating
Policy
0
2.924
Threshold = 3 meter
74.6 (52.5%)
3.05214
Threshold = 2 meter
37 (26%)
2.956
We expected that our updating algorithm would save the communication cost by
more than 70% in the test bed because the track consisted of only three straight lines.
Saving only 52.5% is not satisfactory. The reason of this disappointing test results
stems from the assumption that the Kalman filter makes. The Kalman filter assumes
that the measurements are normal distributed around the real value. This implies that if
we perform our positioning algorithm many (100) times at a point, P, without moving
around, then the average of the measured positions should be P. In the reality, the
average of the positions measured by our positioning algorithm does not exactly
coincide with the real position.
3.3. Rendering Drawings
The purpose of the MODB is to answer spatio-temporal queries. Therefore, most
answers are displayed on a map or a drawing for indoor applications. Our map
rendering module displays a drawing recorded in an AutoCAD DXF (Drawing
Interchange Format or Drawing Exchange Format) file and provides map manipulation
functions such as zoom-in, zoom-out, and move. A DXF file consists of a large number
of entities. The entities are classified into the arc, line, circle, and polyline types,
among others. In DXF, an arc is determined by its attributes such as its center, radius,
start angle, and end angle. The attributes of a line are the start and end points.
Similarly, the attributes of each entity type is predefined in DXF. In DXF, every entity
representation follows the entity type as shown in Table 6. The type of the entity shown
in Table 6 is LINE, and the type name LINE follows “0,” which indicates that the entity
type appears in the next line. The “10” in the table indicates that the X coordinate of the
start point appears in the next line. The others are obvious.
15
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
Table 6. A Part of the DXF File Used in the Proposed Rendering Module
code
…
0
LINE
…
10
35.88
20
13.65
…
11
58.15
21
40.68
…
Description
Indicates the start of an entity
Entity type
The next line is the X coordinate of the start point
X coordinate
The next line is the Y coordinate of the start point
Y coordinate
Z is omitted
The next line is the X coordinate of the end point
X coordinate
The next line is the Y coordinate of the end point
Y coordinate
A lot of entities and others
The main parts of our rendering module are the input, drawing, and manipulation
parts. A flowchart of the input part is shown in Figure 9. The first method is
PaperLoad( ), which uses the dialog tool of C# and returns the full path of the DXF file.
It then continues reading two lines from the DXF file until the end of the file. We read
two lines at a time because the entity type follows “0” in the DXF. The entity type
determines the way of reading the attributes of the entity. Therefore, if the entity type
(for example LINE) is read, we invoke the method (LineModule in this case) which
reads in the attributes of the type. For each entity type, we define a class with local
variables corresponding to the attributes of the entity. It also has methods including
draw( ). For example, x_start, y_start, x_end, and y_end are some of the variables of the
LINE class. For each entity read, our input module creates an object and appends it to
the list called EntityList.
Figure 9. A Flowchart of the Input Part of our Rendering Module
Our drawing part draws all the objects listed in the EntityList. Before the actual
drawing, it calculates the coordinates of the center of the drawing, paperMid and the
16
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
center of the window, windowMid. The center of the drawing is defined by the
following:
paperMidX = (XMin+XMax)/2,
paperMidY = (YMin+YMax)/2,
(Expression 3)
where XMin is the minimum of x coordinates appearing in the entire DXF file.
XMax, YMin, and YMax are all defined similarly. The center of the window is defined
by the following:
windowMidX = drawPanelWidth / 2,
windowMidY = drawPanelHeight / 2,
(Expression 4)
where drawPanelWidth (Height) is the width (height) of the panel where the drawing
is drawn. We now calculate the scale for the drawing, which is basically the ratio of the
window size to the drawing size. It is called mainScale and can be calculated by
(Expression 5). We choose the minimum out of scaleX and scaleY and then multiply it
by 0.9 so that the drawing can easily fit into the window.
mainScale = MIN(scaleX, scaleY) * 0.9, where scaleX = drawPanelWidth / (MaxX –
MinX), scaleY = drawPanelHeight / (MaxY – MinY).
(Expression 5)
Given a point (paperX, paperY) from the DXF file, we can calculate the point on the
window, (winX, winY), by using (Expression 6). A point (paperX, paperY) in the DXF
file is represented at (winX, winY) on the window. In (Expression 6), we use a single
operand negation operation because Y coordinates increase from the top to the bottom
for a window, whereas they increase from the bottom to the top for a drawing. Once we
have the window coordinates, drawing an entity is just a matter of calling the right
method provided by C#.
winX = (paperX – paperMidX) * mainScale + windowMidX,
winY = - (paperY – paperMidY) * mainScale + windowMidY.
(Expression 6)
The manipulation part of the rendering module provides zoom in, zoom out, up,
down, left, right, and so forth. It is easy to implement these functions. For example,
zoom in (out) can be done by increasing (decreasing) the value of mainScale. For
returning to the original scale, the OriginalScale( ) method is provided. Move up
(down) can be done by increasing (decreasing) paperMidY, and move right (left) can be
done by increasing (decreasing) paperMidX. Converting DXF coordinates into window
coordinates can be done by performing (Expression 6). The inverse of (Expression 6)
can be used for converting window coordinates into DXF coordinates.
Experimental results of the rendering program are shown in Figure 10. In the figure,
screen captures after moving, zoom in, zoom out operations are included.
17
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
Figure 10. Experimental Results of the Rendering Module
3.4. Web Service
It is obvious that software reusing improves productivity a lot. Web service is one of
the most convenient ways of reusing software. Therefore, my lab staffs have
transformed those programs introduced in the previous subsections into web methods
for the last year. Uploading location information of APs, returning the distance value
corresponding to the input RSSI value, creating K-NN lookup tables, the real time
phase of the K-NN, trilateration process, uploading RFID tag-ID with its location, and
so on are few out of the implemented web methods. Implementation of these web
methods is shown in Figure 11.
Figure 11. Web Methods for Indoor Positioning
18
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
4. Conclusions
We have worked on indoor location based service for the last a few years and some
of our research results are summarized in this paper. Now, we are planning to
1) Develop practical ILBS systems: integrating the above introduced methods .
2) Improve the methods introduced in this document: developing practical
positioning techniques (for example, utilizing the sensors on the smart phones, figuring
out more efficient algorithms), developing DB of drawings/contents, developing
techniques to generate a drawing/contents by combining many small drawings/contents
published on the Web, practical ontology, and so on.
References
[1] Mabrouk M, “OpenGIS Location Services (OpenLS): Core Services OGC 03-006r”, http://www .opengis.org/
[2] Enge P, Misra P, “Special Issue on GPS: The Global Positioning System”, Proceedings of the IEEE, (1999)
January, pp. 3-172.
[3] Tekinay S, “Special Issue on Wireless Geolcation Systems and Services”, IEEE Communications Magazine,
(1998) April.
[4] Want R, Hopper A, Falcao V, Gibbons J, "The Active Badge Location System", ACM Transactions on
Information Systems, vol. 10, no. 1, (1992) January, pp. 91-102.
[5] Harter A, Hopper A, "A New Location Technique for the Active Office", IEEE Personal Communications,
vol. 4, no. 5, (1997) October, pp. 43-47.
[6] Priyanthat N, Chakraborty A, Balakrishnan H, "The Cricket Location-Support System", Proc. of 6th ACM
International Conference on Mobile Computing and Networking, Boston, MA, (2000) August.
[7] Orr RJ, Abowd GD, “The Smart Floor: A Mechanism for Natural User Identification and Tracking”, In
Conference on Human Factors in Computing Systems (CHI2000) The Hague, Netherlands, (2000) April, pp.
1-6.
[8] Krumm J, et. al., “Multi-camera Multi-person Tracking for Easy Living”, In 3rd IEEE Int‟l Workshop on
Visual Surveillance, Piscataway, NJ, (2000), pp. 3-10.
[9] Bahl P, Padmanabhan V, "RADAR:An in-building RF-based user location and tracking system", INFOCOM
2000, (2000) March, pp. 775-784.
[10] Ladd AM, Bekris KE, Rudys A, Kavraki LE, Wallach DS, Marceau G, “Robotics-based Location Sensing
Using Wireless Ethernet”, Proceedings of the eighth Annual International Conference on Mobile Computing
and Networking (MOBICOM-02), New York, (2002), Sept. 23-28, pp. 227-238.
[11] Yim J, “Introducing a decision tree-based indoor positioning technique”, Expert Systems with Applications,
vol. 34, issue 2, (2008), pp. 1296-1302.
[12] Yim J, Joo J, Park C, “A Kalman Filter Updating Method for the Indoor Moving Object Database”, submitted
for publication.
[13] Wolfson O, "Moving Objects Information Management: The Database Challenge", Proceedings of the 5th
Workshop on Next Generation Information Technologies and Systems (NGITS'2002), Dan Caesarea Hotel,
Caesarea, Israel, (2002) June 25-26, pp. 1-13.
[14] Olston C, Widom J, "Offering a Precision-Performance Tradeoff for Aggregation Queries over Replicated
Data", 26th International Conference on Very Large Data Bases, Cario, Egypt, (2000) September.
[15] Tao Y, Xiao X, Cheng R, "Range Search on Multidimensional Uncertainty Data", ACM Transactions on
Database Systems, vol. 32, Issue 3, (2007) August, pp. 1-54.
[16] Kilimci P, Kalipsiz O, “Moving Objects Databases In Space Applications”, 3rd International Conference on
Recent Advances in Space Technologies, (2007), pp. 106-108.
[17] Ooi B, Huang Z, Lin D, Lu H, Xu L, “Adapting Relational Database Engine to Accommodate Moving
Objects in SpADE”, IEEE 23rd International Conference on Data Engineering, (2007), pp. 1505 – 1506.
[18] Jin P, Wan S, Yue L, “Conceptual Modeling for Moving Objects Database Applications”, 9th International
Conference on Mobile Data Management, (2008), pp. 217–218.
19
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
[19] Guting R, “How to Build Your Own Moving Objects Database System”, International Conference on Mobile
Data Management, (2007), pp.1-2.
[20] Makki S, Sun B, Khojastehpour M, “An efficient information access scheme for mobile objects”, IEEE
International Conference on Information Reuse & Integration, (2009), pp. 312-317.
[21] Han L, Wu J, Xie K, Ma X, Xu D, Zhang H, Chen Z, “A spatio-temporal database prototype for managing
moving objects in GIS”, IEEE International Geoscience and Remote Sensing Symposium.
[22] Zhang W, Li J, Zhang W, “Spatio-temporal Pattern Query Processing based on Effective Trajectory Splitting
Models in Moving Object Database”, First International Multi-Symposiums on Computer and Computational
Sciences, (2006), pp. 540-547.
[23] Kilimci P, Kalipsiz O, “Geometric Modeling and Visualization of Moving Objects on a Digital Map”,
Geometric Modeling and Imaging, (2007), pp. 39-43.
[24] Kuijpers B, Vaisman A, “A Data Model for Moving Objects Supporting Aggregation”, IEEE 23rd
International Conference on Data Engineering Workshop, (2007), pp. 546-554.
[25] Xiao Y, “Set Nearest Neighbor Query for Trajectory of Moving Objects”, Sixth International Conference on
Fuzzy Systems and Knowledge Discovery, (2009), pp. 211-214.
[26] Liao W, Wu X, Zhang Q, Zhong Z, “Improving throughout of continuous k-nearest neighbor queries with
multi-threaded techniques”, IEEE International Conference on Intelligent Computing and Intelligent Systems,
(2009), pp. 438-442.
[27] Iijima Y, Ishikawa Y, “Finding Probabilistic Nearest Neighbors for Query Objects with Imprecise Locations”,
Tenth International Conference on Mobile Data Management: Systems, Services and Middleware, (2009), pp.
52-61.
[28] Ding H, Trajcevski G, Scheuermann P, “Efficient Similarity Join of Large Sets of Moving Object
Trajectories”, 15th International Symposium on Temporal Representation and Reasoning, (2008), pp. 79-87.
[29] Schneider M, “Evaluation of spatio-temporal predicates on moving objects”, 21st International Conference on
Data Engineering, (2005), pp. 516-517.
[30] Xiong X, Elmongui H, Chai X, Aref W, “Place: A Distributed Spatio-Temporal Data Stream Management
System for Moving Objects”, International Conference on Mobile Data Management, (2007), pp. 44-51.
[31] Frihida A, Zheni D, Ghezala H, Claramunt C, “Modeling Trajectories: A Spatio-Temporal Data Type
Approach”, 20th International Workshop on Database and Expert Systems Application, (2009), pp. 447-451.
[32] Ouri Wolfson A, Sistla P, Chamberlain S, Yesha Y, “Updating and Querying Databases that Track Mobile
Units”, invited paper, special issue of the Distributed and Parallel Databases Journal (DAPD) on Mobile
Data Management and Applications, vol. 7, no. 3, (1999), Kluwer Academic Publishers, pp. 257-288.
[33] Li X, Huang M, Bian F, “An Adaptive Dead-Reckoning Updating Policy for Continuous Query”,
International Conference on Artificial Intelligence and Computational Intelligence, (2009), pp. 194-198.
[34] Jeung H, Liu Q, Shen H, Zhou X, “A Hybrid Prediction Model for Moving Objects”, IEEE 24th International
Conference on Data Engineering, (2008), pp. 70-79.
[35] Wang X, Wang S, Wang Z, Lv T, Zhang X, “A New Position Updating Algorithm for Moving Objects”, First
International Multi-Symposiums on Computer and Computational Sciences, (2006), pp. 496-503.
[36] Yu X, Chen Y, Rao F, Liu D, “Filtering location stream in moving object database”, 15th International
Workshop on Database and Expert Systems Applications, (2004), pp. 645-649.
[37] http://code.google.com/intl/ko/apis/maps/documentation/javascript/
[38] Bruntrup R, Edelkamp S, Jabbar S, Scholz B, "Incremental map generation with GPS traces", Proc. of IEEE
Intelligent Transportation Systems, (2005), Sept. 13-15, pp. 574-579.
Authors
Jaegeol Yim received the M.S. and Ph.D. degrees in Computer
Science from the University of Illinois at Chicago, in 1987 and
1990, respectively. He is a Professor in the Department of Computer
Science at Dongguk University at Gyeongju Korea. His current
research interests include Petri net theory and its applications,
Location Based Service, AI systems, and multimedia systems. He
has published more than 50 journal papers, 100 conference papers
20
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
(mostly written in Korean Language), and several undergraduate
textbooks.
Gyeyoung Lee received the M.S. degree in Computer Science from
the Dongguk University in 1982 and the Ph.D. degree in Computer
Engineering from the Dankook University in Korea in 1992, respectively.
He is a Professor in the Department of Computer Science at Dongguk
University at Gyeongju, Korea. His current research interests include
Petri net theory, AI systems and Speech processing systems.
21
International Journal of Grid and Distributed Computing
Vol. 5, No. 1, March, 2012
22
Was this manual useful for you? yes no
Thank you for your participation!

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

Download PDF

advertising