You are on page 1of 6

Enhanced Adaptive Traffic Signal Control System Using Camera Sensor and

Embedded System

2
l
l
l
F. Al Afif I, M. Febrian Rachmadi \ A. Wibowo1 , W. Jatmiko , P. Mursanto , and M. Anwar Ma'sum
IFaculty of Computer Science, 2 Informatics Dept. Faculty of Mathematics and Natural Science
IUniversitas Indonesia, 2Universitas Diponegoro
IDepok, West Java, 2Semarang, Central Java
I, 2Indonesia
Abstract:
Traffic plays an important role in social stability and
community development. Without an appropriate traffic
signal control system, the possibility of traffic congestion
will be very high and causes various negative impacts.
The traffic signal control system with video camera
sensor

is

implemented

in

embedded

systems

using

BeagleBoard-xM. The system uses Viola-Jones method


and Haar Training in detecting the vehicle object from a
video frame. Then, Euclidean distance and kalman filter
methods are used in tracking the vehicle. The ability of
kalman filter in predicting the next position of the object
is a very important feature for multi-object tracking. The
number

of

counted

vehicles

in

each

lane

at

the

intersection then will be processed using Fuzzy Logic to


determine optimal cycle time and split time.
1. INTRODUCTION

Traffic has an important role in daily life. Traffic is the


backbone of capital flows, logistics, and various community
activities. One factor that cause delays in traffic is traffic
congestion. Traffic congestion cause various negative
impacts, including loss of productive time and manpower,
waste of fuel, late delivery of goods, the pollution, and so on.
Traffic congestion in all comers of Jakarta resulting losses to
Rp. 2 8.1 trillion per year [1]. If there is no improvement with
the transportation system until 2020, the estimation of
economic losses would reach Rp. 65 trillion per year [2].
A possible solution to reduce traffic congestion is by
implementing adaptive traffic signal control system. Most of
traffic signal control systems are still a stand-alone system,
where the duration of traffic light is set manually by the
officer. By implementing adaptive traffic signal control
system, the system will be able to adapt to the current traffic
condition. As example, a particular lane with higher traffic
density can have a longer duration of green light.
In adaptive traffic signal control system, the acquisition of
traffic density is a very important issue. Many big cities in
Indonesia currently have a lot of CCTV cameras installed at
the intersections [3], but the camera is merely monitoring the
situation and does not contribute to traffic signal control
system. By using methods in computer vision and image
processing, the camera can be turned into a sensor that can
calculates the number of vehicles for each lane. The method
used to perform this task is Viola-Jones method for vehicle

978-1-4577-1362-0/11/$26.00 2011 IEEE

- 367 -

detection and Kalman Filter method for moving vehicle


tracking. Programs that are made using both methods were
developed using OpenCV library which has been widely
used in both image and video processing.
The number of counted vehicles in each lane at the
intersection then will be processed using Fuzzy Logic. The
fuzzy control was at least as good as the traditional adaptive
traffic signal control for isolated traffic signal. In this model,
the fuzzy logic is used to determine optimal cycle time and
split time.
In line with this situation, some researchers have been
attempting to model a new algorithm to optimize the
performance of traffic signal control system. The aim is to
try to minimize the traveling delay of vehicles through a
network of intersections. Arterial control, such as green
wave [4], is an efficient strategy but only to some arterials by
sacrificing branches benefits. Ehsan Azimirad presents a
novel fuzzy model and a fuzzy logic controller for an
isolated signalized intersection. The controller controls the
traffic light timings and phase sequence to ensure smooth
flow of traffic with minimal waiting time and length of queue
[5]. Hong used fuzzy control to manage the urban traffic
signal [6] and Bingham built the model of urban traffic
control using the fuzzy-neuro method [7]. Most of works are
focus on single or stand alone controller algorithm.
Traffic signal control system developed by the Faculty of
Computer Science Universitas Indonesia currently has
reached the stage of prototype performance analysis, where
the performance of each component in the system is
analyzed in the term of accuracy as well as connectivity with
other components. The prototype uses BeagleBoard-xM as
the primary engine, web camera as a replacement of a CCTV
camera, and AVR microcontroller minimum system.as the
traffic light display. Figure 1 shows architecture diagram of
traffic signal control system.

Figure 1. Architecture Diagram of Traffic Signal Control System.

2. PREVIOUS WORK

The aim of this study is to create a traffic data acquisition


tool and adaptive traffic signal control system. Adaptive
traffic signal control system is intended to meet the need of
an adjustable setting of traffic signal control so that it can be
adaptive to the current traffic conditions. In previous studies
we have developed an adaptive traffic signal control system
with vehicle detection and counting using the blob tracking
method and Principal Component Analysis (PCA) [ 8].
However, blob tracking method requires considerable
computation process and memory resources, whereas the
algorithm means to be implemented on the embedded
systems device which has a limited computation speed and
available memory. So, we try to find alternative algorithms
that are lighter but more powerful. Therefore we have
developed traffic data acquisition tool using the PCA and
euclidean distance as a method for obtaining the information
of traffic density and using the Distributed Constraint
Satisfaction Problem (DCSP) as a traffic signal control
algorithm. DCSP determines the signal duration of traffic
lights at each lane of the intersection. The decision making
on the prototype at this point is still a comparison of many
vehicles in each lane, yet pure DCSP. The development of
simulation software of DCSP-based traffic signal control
system is still in progress [9] and is performed on different
team with the development of this traffic signal control
system prototype.
We have implemented the above methods on the
BeagleBoard-xM embedded systems with several supporting
devices. The algorithms are quite light to be run on the
embedded system device but the tracking problem and some
errors in detection can not be accommodated by the
algorithms, so that the accuracy of the vehicle detection and
counting is yet not at the optimum level.
3. METHOD
3. 1 Vehicle detection using Viola-Jones method

Viola-Jones method published by Paul Viola and Michael


Jones in 2001 and is now often used to detect objects quickly
in images or video. This method to detecting objects in
images combines four key concepts as follow [10].
1. Simple rectangular features, called Haar features
2. An integral Image for rapid feature detection
3. The AdaBoost machine-learning method
4. A cascaded classifier to combine many features
efficiently
The features used by Viola-Jones method are the features
based on Haar wavelets [10]. Haar wavelets are the single
square waves (one high and one low interval). In the
two-dimensional image, a square wave is formed from a pair
of adjacent rectangles, one light and one dark. In its
implementation, the actual rectangle combinations used for

- 368 -

visual object detection are not true Haar wavlets, but rather a
combination of rectangles that better suited to visual
recognition tasks. Therefore, these features used in the
Viola-Jones method are called Haar features, or Haarlike
features, rather than Haar wavelets. In OpenCV library that
is used by this study, the Viola-Jones method was developed
to detect objects quickly, for example, is a human face.
Figure 2 shows an example of Haarlike features used in the
Viola-Jones method in OpenCV.

Figure 2. Examples of the Haar features used in OpenCV


(SERVO Magazine, February 2007).
3.2 Haar Training

Haar training is a training process that is used to detect


complex objects in images or video streams. Haar training is
one approach to training with a statistical model using a
series of negative images, i.e. images that do not contain
objects to be identified (e.g. background or other objects that
are not relevant) and a series of positive images, i.e. images
that contain objects to be identified (in this study is a
vehicle). The aim is to extract the distinctive features and
characteristics of an object. Haar training will produce a file
containing the typical features of an image object called as
classifiers [11 ]. Haar training usually used to detect specific
objects, for example the detection of human faces.
Haar training is one of the four main concepts in the
Viola-Jones method because Haar training use machine
learning technique Adaptive Boosting (often abbreviated and
referred as AdaBoost) in extracting Haarlike features. More
specifically, the Haar training association with the OpenCV
is already well known because this library is a good
implementation of Haar training using AdaBoost machine
learning techniques for Viola-Jones detection methods.
There are at least four types of images needed to perform
the haar training as follow [12 ].
1. Positive images, i.e. images that only contain objects
that want to be trained.
2. Negative image, i.e. a background image that there is
absolutely no positive object in it.
3. Examples of positive images, the background image
that also contains a positive object in it.
4. Examples of negative images, the background image
that there is absolutely no positive object in it. Examples of
negative images may be the same to the negative image, but
it is recommended that we use a different'

Figure 3. Example of positive images of the vehicle.

3.3 Object Tracking using Kalman Filter Method

The purpose of Kalman Filter is to take measurements


observed from time to time where the measured data contain
noise (random variation) and other inaccuracies. Kalman
filter will produce values that are close to the actual value
and other calculated values of mutual association. Kalman
filter is a recursive solution that is used to solve problems of
linear data separation in discrete time [13].
Kalman filter consists of two main processes, namely the
"prediction" and the "correction" carried out repeatedly on a
moving object (Figure 4). The process of prediction requires
the position of objects at an earlier time to predict the
position of the object at the moment. The process of
correction requires a measurement result of the actual
position of the object at the moment. The process of
"correction" will correct the predicted position of the object
during the process of "prediction", so the predicted position
of the object will approach the position of the actual object.
The features of the Kalman filter that can predicts the next
position of the object according to the previous object
position is a very important feature for tracking objects,
including for multi-object tracking.

". ,

Change of lime Change ofposilion


"Correction"

"Prediction"

advantage unlike other optimization settings, that is not rely


on mathematical modeling functions to formulate control
strategies. In contrast, a fuzzy signal controller design
requires expert knowledge and experience of traffic control
in formulating linguistic protocol, which in tum generates
control inputs to be applied to the traffic signal system [14].
The use of fuzzy in the traffic signal control was first
made by Pappis and Mamdani in 1977, which made the
study of theoretical simulation of the fuzzy logic controller
in an isolated/independent intersection (2 +2 lanes, one-way
intersection) [15]. In their study, Pappis and Mamdani was
comparing fuzzy method with adaptive traffic signal control
to reduce vehicle delays by calculating the optimal cycle
time. The research shows that the fuzzy controller gives
results that are similar or slightly better than the adaptive
method. The basic idea of the fuzzy set theory is as a
quantitative interpretation guide or in other words, is for
decision-making process of the cycle time and split time in
the traffic signal control. Fuzzy control instructions in the
traffic signal control is generally in the form [16]:
IF Cycle Time
AND Arrivals (Vehicle Arrival Rate)
AND Queue (Queue Vehicles)
THEN Extension (Addition Time of Cycle Time)
The study and also the research on optimization of fuzzy
control in traffic signal control continues in making progress
until now [5, 6, 7, 14, 16, 17].
4. IMPLEMENTATION

Figure 4. Turnaround process "Correction" and the "prediction" in


accordance with change of time and position of the object.

4. 1 Traffic video processing in vehicle counting

3.4 Fuzzy Logic in Traffic Signal Control System

Fuzzy Logic developed in traffic signal control system is


referring to the concept of human decision-making process.
The aim is to simplifY the decision-making process by
referring to human concept in the information acquisition
and then softly executing based on decision rule. Fuzzy
control has been developed in the context of fuzzy inference.
Fuzzy inference is the process of inference based on a
multi-valued logic inference, in other words, the truth values
of inputs and rules of inference process is not a single value
(yes or no), but a multi-value. As a result, the correctness of
the decision was given a value between 0 and 1, which is
more similar to the conclusions made by human
decision-making process, which involves a greater degree of
ambivalence or less [14].
Fuzzy logic allows linguistic and inexact data to be
manipulated as a useful tool in designing the timing signal. It
also provides a means to convert the linguistic control
strategy, which is expressed by the statement if-then-else,
into a control algorithm. The motivation in designing a fuzzy
controller is that there is a fairly direct relationship between
linguistic expressions apart from the traffic control strategies
and the manual implementation. Fuzzy logic has an

- 369 -

Traffic video processing for counting the number of


vehicles is divided into three stages, namely the input stage,
the stage of data processing, and output stage. All the three
stages should be done so that the processing of traffic video
can run smoothly and produce the desired output, an
accurate counting of vehicles. Figure 5 shows the overview
of traffic video processing in counting the number of
vehicles.
Frame

Video

Rate

Resolu

Object

Video

Normali

tion

Detection

Frames

zation

Normali
zation

Object

Kalman

Object

Counting

Filter

Tracking

Figure 5. Traffic video processing in vehicle counting.

4. 1. 1 Input stage

Input stage is the stage where the camera starts to capture


traffic video until the processing of raw video data into data
that are ready to be processed at a later stage (normalization).
Normalizations performed on the video are resolution and
frame rate settings. Each traffic video will be set to the
resolution of 320 x 240 pixels. Video frame rate will be set
so that it has a frame rate of 30 fps.
4. 1.2 Data processing stage

Data processing stage is the longest stage in the


processing of traffic video. This stage will performs vehicle
object recognition and tracking in every frame of the video.
Moreover, this stage will also perform the process of vehicle
object counting with the help of a virtual gate.

is a necessary, so it can be concluded that the size of state


vector required is 4 x 1 as shown in equation (1). The next
step is to determine the matrix vector measurement.
Measurement vector describing the components to be
predicted in the Kalman Filter. Each object in the traffic
video will be predicted its location on a 2D cartesian field,
so it can be concluded that the size of measurement vector
required is 2 x 1 as shown in equation (2).
The size of state vector will determine the magnitude of
transition matrix used in the Kalman Filter. Since the size of
state vector is 4 xl, then the size of transition matrix is 4 x 4
as shown in equation (3). The size of the measurement vector
will determines the magnitude of the measurement matrix.
Since the size of measurement vector is 2 x 1, then the size
of transition matrix is 2 x 4 as shown in equation (4).
x

4. 1.2. 1 Vehicle Object Detection

Vehicle object detection is performed on each frame of the


traffic video using Viola-Jones method with classifier data
resulted from Haar training process. The accuracy of object
detection depends on the Haar training. Higher accuracy of
object detection would be achieved by using more positive
data in the training and error rate of object detection would
be achieved by using more negative in the training.
Object detection using the Viola-Jones method can
produce several objects ranging from 0 up to n objects, or in
other words, will produce a multi-object. Each of these
objects will undergo into the process of object tracking using
Kalman Filter. The data structure used in this research is
linked-list data structure. To determine whether the object
was already in the linked list as a result of object detection in
the previous frame, we need to do data association checking
using euclidean distance method.
4. 1.2.2 Object Tracking

Two objects on the different frame is determined as the


same object by calculating the distance of two center points
of both objects using the euclidean distance method. The
closer distance of two center points, the higher possibility
that two objects are actually the same object. Distance
difference tolerance is set to 30 pixels but can be changed as
needed during implementation.
The next important component of object tracking is to
estimate the position of objects on the next frame. In image
processing techniques, there are various methods that can be
used for this purpose. One well known method is Kalman
Filter method. The advantage of Kalman filter is its ability to
predict not only distance, but also speed or even acceleration
of a moving object. The more things to be predicted
(distance, speed, and so on) and the higher the dimension
(2D, 3D, and so on), the greater computational costs needed.
The position of the object (referred as states) on the
Kalman Filter is a vector that will be operated on the
functions of Kalman Filter. Vehicle object on the video is
within a field of 2D Cartesian and speed change on that field

- 370 -

state vector

dx/
dt
d y/

dt

measurement vector

tT=ition matT'"

(1)

[i

measurement matrIX

G]

(2)

dt

1
0

[1

0
1

dt

']

(3)

1 ]

(4)

Where:
x= x coordinates of cartesian field.
y= y coordinates of cartesian field.
dx/dt = object speed with respect to time on x cartesian field.
dy/dt = object speed with respect to time on y cartesian field.
dt = change of the time between successive frames.
In developing the program, we use OpenCV because
OpenCV Kalman Filter has been providing functions for
both subprocess. The prediction prosess is performed on
each object in the linked list if there is a change of position
of the object. The correction prosess is performed on each
object in the linked list if there is a change of time, or in
other words, the process is done on every changes in the
frame of the traffic video.
4. 1.2.3 Vehicle Counting

If the performance of object detection and tracking in the


system is already fine, then the object is easier to be counted
as the error in counting will be minimized by itself.
However, if the performance is still depending on the
fluctuating environment conditions, additional method will
be needed to maximize the accuracy of the object counting.
The method used in this study is a virtual gate method.
Virtual gate or a Region of Interest (ROO on the frame
image is used as an area or border of the vehicle detection
and tracking. If the vehicle detection and tracking stopped

Rule:

before the virtual gate, then the vehicle will be counted as


one vehicle. However, if it stopped outside the virtual gate,
then the vehicle will not be counted. The definition of
stopped detection and tracking is that the vehicle no longer
detected and tracked for more than 150 times, which means
the object is never detected again for the 150 next frames.
4. 1.3 Output stage

At the output stage, the result of object detection and


tracking is a counted number of vehicles that are constantly
updated and displayed during the running of the program.
4.2 Fuzzy Logic Implementation

Implementation of fuzzy logic in traffic signal control


system is by applying the basic principles of fuzzy settings
that have been proposed by Mamdani [15]. The
defuzzyfication used in this method is the center of gravity
(COG) which produces the best performance in a stable
condition and without extraordinary traffic events that may
not occur in a simulation [15]. We use fuzzy lite 1.03
developed in c programming language as the fuzzy logic
library [1 8]. The fuzzy inference is as follows. There are 3
input variables used in fuzzy inference and there is 1 output
variable as a result of fuzzy inference. The results of fuzzy
inference is as a setting for traffic signal control based on
video camera input from the traffic information previously
obtained from the system.
Input variables:

10

15

20

25

5. 1 Accuracy of Vehicle Detection and Counting

The vehicle detection and counting program developed by


Faculty of Computer Science Universitas Indonesia has
experienced several developments using many different
methods of image processing techniques. The previous
version of program only uses the Euclidean Distance method
without Kalman Filter. Thus the Euclidean Distance is used
as a method of object tracking as well as vehicle data
association checking between frames. The addition of
Kalman filter method as a tracking method and estimation of
the vehicle position in the next frame is expected to improve
the accuracy of the vehicle counting. Table 1 and table 2
show the accuracy of vehicle detection and counting
program on a variety of weather and traffic conditions.
Figure 6 shows the vehicle detection and counting
accuracy of the previous and latest version of the program.
The average level of counting accuracy can be increased by
2.77%, with the highest increase is 8. 62% (cloudy day
normal traffic condition). The increase of accuracy can also
be achieved for sunny day traffic jam condition, in the
amount of 3.45%, which makes the system achieved 100%
of accuracy. The results show that the addition of a Kalman
filter as an object tracking method gives good results
because it can increase the accuracy of vehicle tracking.

30

2.

3.

Green time (seconds)

EZSXd
o

10

20

30

Table l. The accuracy of the previous vehicle detection and counting


program on a variety of weather and traffic conditions.
Weather and Traffic
Condition

Output variables:
I.
G r een time extension
MI us

STATE

Sunny Day in Traffic


Jam
Cloudy Day in
Normal Traffic
Sunny Day in
Crowded Traffic (I)
Sunny Day in
Crowded Traffic (2)

PI.

0.5

10

-8

-6

-4

-2

oup,t veri8ble "greenTmef:denslon"

if density is LOW and Velocity is SLOW and greenTime is


SHORT then greenTimeExtension is PLUS
if Density is LOW and Velocity is SLOW and greenTime is
MEDIUM then greenTimeExtension is STATE
if Density is LOW and Velocity is SLOW and greenTime is
LONG then greenTimeExtension is MINUS
if Density is MEDIUM and Velocity is SLOW and greenTime
is SHORT then greenTimeExtension is PLUS
if Density is MEDIUM and Velocity is SLOW and greenTime
is MEDIUM then greenTimeExtension is PLUS
if Density is MEDIUM and Velocity is SLOW and greenTime
is MEDIUM then greenTimeExtension is MINUS
if Density is HIGH and Velocity is SLOW and greenTime is
SHORT then greenTimeExtension is PLUS
if Density is HIGH and Velocity is SLOW and greenTime is
MEDIUM then greenTimeExtension is PLUS
if Density is HIGH and Velocity is SLOW and greenTime is
LONG then greenTimeExtension is STATE
5. EXPERIMENT AND RESULT

L 1SZX1
o

10

371

Counted
Vehicle
by
Manual
(M)

PCA&
Eucli
dean
Distance
(PEl

Error
I MPEl

87

84

3.45

96. 55

58

42

16

27. 59

72.41

102

109

6.86

93.14

128

120

6.25

93.75

Error

(%)

Accur
acy

('Yo)

Table 2. The accuracy of the latest vehicle detection and counting program
on a variety of weather and traffic conditions
Weather and Traffic
Condition

REFERENCES

Counted
Vehicle
by
Manual
(M)

ViolaJones&
Kalman
Filter
(VK)

Error
I MVK I

87

87

0.00

100.0

[2]

58

47

II

18.97

81.03

[3]

102

94

7.84

92.16

[4]

128

120

6.25

93.75

Sunny Day in
Traffic Jam
Cloudy Day in
Normal Traffic
Sunny Day in
Crowded Traffic (I)
Sunny Day in
Crowded Traffic (2)

Error

(%)

Accur
acy

[I]

(%)

indexphp? optioncomJontent&viewarticle&id 153:infrastruktur

[5]

Vehicle Detection and Counting Accuracy


10 000%
9500%
9000%

>:
"
E

<f!.

85.00%

::J

80.00%

7500%

"
"

[6]

C'\.
\."\..
/
\."\,..
.4
\. '\/ /
\. /

[7]
[8]

70.00%

[9]

6500%
6000%

At Cloudy Day

At Sunny Day

At Sunny Day

in Normal

in Crowded

in Crowded

Traffic

Traffic(1)

Traffic(2)

96. 55%

72.41%

93.14%

93.75%

100.00%

81.0 3%

92.1 6%

93.75%

At Sunny Day
in Traffic Jam
Previous
Version
Latest
Version

[10]

Variety of weather and traffle conditions

Figure 6. Comparison of vehicle detection and counting accuracy of the


previous and latest version of the program.
6. CONCLUSION

[II]

[12]

The results of vehicle detection using the Viola-Jones


method deliver optimal results with sufficient training data.
Vehicle tracking using Kalman filter method provides an
accurate tracking of vehicle by evaluating the detection in
every frame. Kalman filter as object tracking method gives
an increase of the overall accuracy. Implementation of
vehicle detection and tracking on BeagleBoard shows that
the used method is light and reliable for further
optimizations. Implementation of traffic signal control using
fuzzy logic provides a setting that is more adaptive and has
been proven to be optimal in other publications. In further
studies the system will be tested on other traffic conditions
or by evaluated on real traffic simulation to verify the traffic
signal control system performance.
7. ACKNOWLEDGEMENT

This research is supported by Incentive Research Program


and Grant from Minister of Research and Technology
Republic of Indonesia and Hibah Riset Unggulan Universitas
Indonesia 2011.

- 372 -

Ali, Firdaus. "Transportasi Massal Versus Kemacetan". BAPPENAS


Directorate of Transportation. hUp://transportasi.bappenas.go.id/

[13]
[14]

[15]

[16]

[17]
[18]

s1
&catid56:berita-kegiatan&Itemid55. Last accessed on March, 21
2011.
Yayasan Pelangi. http://wwwpelangi.or.id/pressphp?persid64 .
Last accessed on December, 26th 2010.
Traffic Management Center. http://www.tmcmetro.com/. Last
accessed on December, 26th 2010.
N. H.Gartner, S. F.Assmann, F. Lasaga, and D. L. Hom. A multiband
approach to arterial traffic signal optimization. Transport. Res. B, vol.
25, pp. 55-74, 1991.
Azimirad, E., Pariz, N., & Sistani, M. B. A Novel Fuzzy Model and
Control of Single Intersection at Urban Traffic Network. IEEE
SYSTEMS JOURNAL, Vol. 4. No.I,107-111. 2010.
W. Hong and x.-g. Mu. A cooperative fuzzy control method for
traffic lights. J. Syst. Simul., vol. 13, no. 5, pp. 551-553, 2001.
E. Bingham. Reinforcement learning in neurofuzzy traffic signal
control. Eur. J. Oper. Res., vol. 131, pp. 232-241, 2002.
Jatmiko, W., et.a!. "Detection and Counting of Vehicles based on
Video Processing in Distributed Traffic System", International
Conference on Advance Computer Science and Information System
2010.
Jatmiko, W., Hariyadi, F., Krisnadhi, A A, Takagawa, I., Sekiyama,
K., & Fukuda, T. Distributed Traffic Control with Swarm-Self
Organizing Map in Jakarta: Simulation and Measurement.
International Symposium
on Micro-Nano Mechantronics and
Human Science, MHS, 598-601. 2009.
Hewitt, Robin. 2007. "How Face Detection Works". Cognitics,
Resources for Cognitive Robotics. hup://www.cognotics.com/opencv/
servo_2007_series/part_2/sidebar.html. Last accessed on August,
24th 2011.
Kuranov, Alexander, Rainer Lienhart, dan Vadim Pisarevsky. "An
Empirical Analysis of Boosting Algorithms for Rapid Objects With
an Extended Set of Haar-like Features". Intel Technical Report
MRL-TR-July02-01, 2002.
Seo, Naotoshi. 2008. "Rapid Object Detection With A Cascade of
Boosted Classifiers Based on Haar-like Features". Tutorial: OpenCv'
hUp://note.sonots.com/SciSojtware/ haartraining.html. Last accessed
on March, 21't 20II.
G. Welch and G. Bishop. "An Introduction to the Kalman Filter,"
UNC-Chapel Hill, 2004.
Kim S. Applications of Petri Networks and Fuzzy Logic to Advanced
Traffic Management Systems. Ph.D. thesis, Polytechnic University,
New York, USA 139 p. 1994.
Pappis C., Mamdani E. A fuzzy logic controller for a traffic junction.
IEEE transactions on systems, man and cybernetics. Vol. SMC-7, No.
10, pp. 707-717.1977.
Niittymiiki, Jarkko. Fuzzy Traffic Signal Control: Principles And
Applications. Helsinki University of Technology Transportation
Engineering Publication. 2002.
Slinn, M., Paul Matthews, Peter Guest. Traffic Engineering Design,
2nd Edition . Elsevier Butterworth-Heinemann, Burlington, 2005.
Rada-Vilela, Juan, Sergio Guadarrama, Luis Magdalena. FuzzyLite, A
Fuzzy Inference System written in C++. September 2010.

You might also like