BAYES-BASED OBJECT TRACKING BOOSTED BY PARTICLE

SWARM OPTIMIZATION

Yuhua Zheng and Yan Meng

Department of Electrical and Computer Engineering, Stevens Institute of Technology, Hoboken, NJ 07030, USA

Keywords: Vision, object detection, tr

acking, particle swarm optimization, Bayes law.

Abstract: This paper presents a novel Bayes-based object tracking framework boosted by a particle swarm

optimization (PSO) algorithm, which is a population based searching algorithm. Basically two searching

steps are conducted in this method. First, the object model is projected into a high-dimensional feature space,

and a PSO algorithm is applied to search over this high-dimensional space and converge to some global

optima, which are well-matched candidates in terms of object features. Second, a Bayes-based filter is used

to identify the one with the highest possibility among these candidates under the constraint of object motion

estimation. The proposed algorithm considers not only the object features but also the object motion

estimation to speed up the searching procedure. Experimental results demonstrate that the proposed method

is efficient and robust in object tracking.

1 INTRODUCTION

Object detection and tracking in images is an active

research area which has attracted extensive attentions

from multi-disciplinary fields, and it has wide

applications in many fields like service robots,

surveillance systems, public security systems, and

virtual reality interfaces. Detection and tracking of

moving object like car and walking people are more

concerned, especially flexible and robust tracking

algorithms under dynamic environments, where

lightening condition may change and occlusions may

happen.

Up to now, the underlying mathematical models

most tracking methods are Bayes’ law estimation

and Hidden Markov Model (HMM). The most

popular approaches to predict discrete probability

distribution are Kalman filter (G. Welch and G.

Bishop, 2001), condensation (M. Isard, 1998),

particle filter (S. Maskell and N. Gordon, 2001) and

mean shift (D. Comaniciu, and P. Meer, 2002).

Kalman filter has the same idea with HMM, while

Kalman filter deals with discrete variables. Some

researchers proposed different control and noise

models into the recursion function for image

processing, however those assumptions are

dependent on varied applications and need to be

tuned carefully. Condensation methods mainly focus

on how to sample probabilities and likelihoods.

When these methods are applied to multiple objects,

a dominant peak is established if an object has large

likelihood values more frequently, which may

depress and lose other objects. The performance of

particle filter based methods is limited by

dimensionality of state space, which may be feasible

in the cases with fewer targets, but may be intractable

with a large amount of targets. Generally speaking,

the mean-shift algorithm is efficient for object

tracking. However the searching window may drift

away from the object under dynamic conditions. For

example, if the kernel is lost from the tracked target

in one frame under some emergent situations, such as

illumination condition change, it would be difficult

for the tracker to recover itself from this unpredicted

event.

Usually for object tracking, an analysis window

ased on the expectation of objects features is built

and scan over the image to find out areas of interest

(AOI). However, most conventional analysis-window

based trackers are influenced by the shape and size of

the window, which may vary from one frame to

another. It is difficult to find the appropriate window

for each frame, especially under dynamic

environments where the content of the images may

be dramatically changed.

There are various features can be used for object

det

ection and tracking, such as color, shape, texture,

101

Zheng Y. and Meng Y. (2007).

BAYES-BASED OBJECT TRACKING BOOSTED BY PARTICLE SWARM OPTIMIZATION.

In Proceedings of the Fourth International Conference on Informatics in Control, Automation and Robotics, pages 101-108

DOI: 10.5220/0001646901010108

 SciTePress

gesture, contour, and motion. Some successful

methods take advantage of knowledge of objects,

such as shape or structures. However, the shape-

based methods cannot handle the cases with

occlusions efficiently. Appearance histogram is

applied as tracking cue in this paper due to its

independency with objects’ shape and structure.

A Bayes-based object tracking approach using a

particle swarm optimization (PSO) algorithm is

employed to search for an optimal window in a super

feature space based on appearance histogram instead

of image plane directly. The PSO algorithm (J.

Kennedy, R. C. Eberhart, and Y. Shi, 2001) was

inspired by the social behavior of a flock of birds. In

PSO algorithm, birds in a flock are symbolically

represented as particles. These particles can be

considered as simple agents “flying” through a

problem space. A particle’s location in the multi-

dimensional problem space represents one solution

for the problem. When a particle moves to a new

location, a different solution is generated. This

solution is evaluated by a fitness function that

provides a quantitative value of the solution’s utility.

The PSO algorithm is effective for optimization

of a wide range of searching problems. In this

problem, particles fly around the feature space, trying

to find the best-fit tracking window parameters based

on the fitness function of object features using

appearance histogram. When some particles

successfully detect the objects, they will share that

information with their neighbors, and their neighbors

may follow the directions to reach objects very

quickly. Each particle makes its own decision not

only based on its neighbors, but also on its own

cognition, which provides the flexibility and ability

of exploring new areas. This decision-making

procedure can efficiently prevent the local optimum,

which may cause the searching window drift.

By using PSO, the problem of identifying

tracking window is translated from one-to-one

estimation into one-to-many searching, which brings

more flexibility. Since this searching procedure is

conducted only in the object feature space, to

improve searching results, a Bayes law filter is

constructed based on the motion constraints of

tracked objects to identify the most possible solution.

Generally it is reasonable to assume that objects

move consecutively in successive frames. The Bayes

law filter tries to keep inertia of the object motion.

Compared with conventional window-tracking

algorithms, the proposed method can be executed

automatically, and moving objects can be detected

and tracked in a more flexible and robust way.

This paper is organized as follows. Section 2

simply reviews some related work in object detection

and tracking. Section 3 introduces the PSO algorithm.

The Bayes-based adaptive-window approach boosted

by the PSO algorithm is described in Section 4.

Experimental results are discussed and analyzed in

Section 5. Conclusion and further work are given in

section 6.

2 RELATED WORKS

There are many systems proposed in the past few

decades for object detection and tracking. Zhang et al.

(Zhang et al., 2006) proposed a robust method to

detect moving objects at distance using a mobile

camera. Through the utilization of the focus of

expansion (FOE) and its associated residual map, the

proposed method is able to detect and separate

independently moving objects (IMOs) from the

"moving" background caused by the camera motion.

Leykin and Hammoud (Leykin and Hammoud, 2006)

used a combined input from RGB and thermal

cameras to build background model and tracker for

pedestrians. This method showed robustness for

outdoor environments. Olson and Brill (T. Olson and

F. Brill, 1997) built a general purpose system for

moving object detection and event recognition,

where objects were detected and tracked by both

first-order prediction and nearest neighbor matching.

The work which is most related to our method is

(Yuri Owechko, Swarup Medasani, and Narayan

Srinivasa, 2004), where the authors treated every

particle as a classifier with different parameters.

Those classifiers swarm in the solution space to

converge to the optimal analysis window. However

this is a simple application of PSO for people

detection only. Reza Akbari etc. (Reza Akbari,

Mohammad Davarpanah Jazi, and Maziar Palhang,

2006) employed both PSO algorithm and Kalman

filter in a hybrid framework of region and object

tracking, where vehicles were tracked in a cluttered

background. A PSO algorithm was proposed in (Luis

Anton-Canalis, Mario Hernandez-Tejera, and Elena

Sanchez-Nielsen etc., 2006) to drive particles flying

over image pixels directly, where object tracking

emerged from interaction between particles and their

environment.

ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics

102

3 PARTICLE SWARM

OPTIMIZATION

PSO algorithm is an efficient optimization method

proposed by Kennedy and Eberhart in 1995 (R.

Eberhart and J. Kennedy, 1995) (J. Kennedy and R.C.

Eberhart,1995) from the simulation of a simplified

social model, which obviously has its root in

artificial life in general, and in bird flocking, fish

schooling and swarming theory in particular. On the

other hand, it is also a method of evolutionary

computation, related with both genetic algorithm and

evolutionary programming.

The PSO algorithm is population-based: a set of

potential solutions evolves to approach a convenient

solution for a problem. Being an optimization

method, the aim is to find the global optimum of a

real-valued fitness function defined in a given search

space. Rather than just being a social simulation,

PSO can be treated as a powerful new search

algorithm, capable of optimizing a wide range of N-

dimensional problems.

The social metaphor that leads to this algorithm

can be summarized as follows: the individuals that

are part of a society hold an opinion that is part of a

"belief space" (the search space) shared by

neighboring individuals. Individuals may modify this

"opinion state" based on three factors:

 The knowledge of the environment (inertia

part)

 The individual's previous history of states

(individual part)

 The previous history of states of the

individual's neighborhood (social part)

An individual's neighborhood may be defined in

several ways, configuring somehow the "social

network" of the individuals. Following certain rules

of interaction, the individuals in the population adapt

their scheme of belief to the ones that are more

successful among their social network. Over the time,

a culture arises, in which the individuals hold

opinions that are closely related.

In the PSO algorithm each individual is called a

"particle", and is subject to a movement in a

multidimensional space that represents the belief

space. Particles have memory, thus retaining part of

their previous states. There is no restriction for

particles to share the same point in belief space, but

in any case their individuality is preserved. Each

particle's movement is the composition of an initial

random velocity and two randomly weighted

influences: individuality, the tendency to return to the

particle's best previous position, and sociality, the

tendency to move towards the neighborhood's best

previous position.

The velocity and position of the particle at any

iteration is updated based on the following equations:

)()(

2211

1 t

xpcxpcvwv −⋅⋅+−⋅⋅+⋅=

ϕϕ

(1)

11 ++

vxx

(2)

where is the component in dimension d of the

ith particle velocity in iteration t, is the

component in dimension d of the ith particle position

in iteration t, are constant weight factors, is

the best position achieved by particle i, is the

best position found by the neighbors of particle i,

,cc

are random factors in the (0,1) interval, and

is the inertia weight. The PSO requires tuning of

some parameters: the individual and sociality

weights , and the inertia factor .

,cc w

The mechanism of PSO implicitly assumes that in

most real world situations, the optima have better

residence around them. Experimentally during the

search, regions with high fitness values attract more

particles and make particles concentrated after a few

iterations. So this type of search is faster and more

effective than traditional scanning and gradient

methods. On the other hand, PSO is simpler than

genetic algorithm since all particles employ the same

mechanism during evolutions. Although basic PSO is

designed for only single optimum, there are many

works have been done to process more complex

issues (Kennedy, J. &R.Eberhart, 1997).

4 THE APPROACH

4.1 General Idea

Basically, object tracking can be considered as a

probability-based classification and estimation,

which searches for the best match of the target model.

Usually searching algorithms rely on two factors:

searching space and searching window. In terms of

the searching space, the more features the object has,

the larger the searching space will be. To expedite

the search, we can either bound the searching space

with some constraints, or develop an efficient

searching algorithm. Considering the searching

window, adaptive windows have been extensively

utilized due to its robustness.

In this paper, we propose a framework which

combines a PSO-based searching algorithm and a

Bayes-based probability algorithm to achieve the

BAYES-BASED OBJECT TRACKING BOOSTED BY PARTICLE SWARM OPTIMIZATION

103

efficiency and robustness of the tracking systems.

Basically, a PSO-based searching algorithm

identifies the changes in the scene, and the

probability-based algorithm estimates the best

candidate of the object with the highest possibility.

More specifically, the PSO algorithm takes fast

scouting in a high-dimensional feature space and

finds out some object candidates. Then Bayes law

filter decides which one is the best match.

4.2 Object Detection

Usually object detection and recognition depend on

the features of the object, such as color, texture, and

shape. As indicated in (J. R. Jain and A. K. Jain,

1981), most changes in video content are typically

due to the motion of objects in the depicted scene

relative to the imaging plane, and a small amount of

motion can result in large differences in the values of

the samples in a picture, especially near the edges of

objects. Often, predicting an area of the current

picture from a region of the previous picture that is

displayed by a few samples in spatial location can

significantly reduce the need for a refining difference

approximation. We call this special displacement

motion vectors.

Since only the moving objects are considered to

be tracked in this paper, the object detection turns

into motion detection where a simple background

subtraction method is applied. When the detection

starts, the first several frames are looked as the

background. In the following frames, the moving

targets can be easily detected by a motion detection

algorithm using background subtraction. During this

procedure, the histogram model of background is

built and updated by averaging every coming frame

to achieve higher robustness. The motion vector

can be obtained, where represents

motion vectors of particle i, and N represents the total

number of particles. Once a valid object is identified,

the tracking algorithm kicks in.

NiV

,...,2,1, =

4.3 PSO-based Searching Algorithm

From the view of searching, the PSO algorithm is a

distributed convergence method. The key is to take

advantage of sharing information between the

particles as well as their own past experiences to

accelerate the convergence. The PSO algorithm

would provide an optimal or near-optimal solution

using appropriate fitness functions without the

complete knowledge of the searching space.

To identify an object in an image, usually

rectangle windows are utilized in most cases. Four

parameters will be identified to describe the rectangle

windows, including 2D location of the central point,

width and height of the rectangle, as shown in Figure

1. These parameters can build up a four-dimensional

search space.

Figure 1: The four parameters associated with a particle

window.

So in such a space, each particle presents a search

window with specific values of parameters, which

can be defined as:

},...,2,1),,,,(|{ NiwlyxppP

iiiiii

(3)

Where represent the central point of the

rectangle related to particle i; and represents

the length and width related to particle i; and N is the

population of swarm particles. Each individual

particle has different values of these parameters. In

other words, they are distributed in a four-

dimensional search space.

yx and

Generally a four-dimensional feature space is

very large, which makes search algorithms to be

computation extensive. Some motion-based

constraints can be applied to limit the search area to a

smaller region where particles are initialized and

move around. A straightforward constraint is using

the continuity of movement since it is reasonable to

assume that motion is continuous under most

tracking situations. In other words, the tracking

window of a new frame should be adjacent to its

previous one. In this way, the initialization of

particles could be accelerated.

Suppose

),,,,(

'''''' bbbbbb

wlyxp

is the best particle

(i.e., tracking window) in last frame, the initialized

particles

),,,,(

iiiiii

wlyxp

, where i = 1,2,…,N, in

the new frame should be around with some

offsets in each dimension. In our experiments,

locations are shifted up to 15 pixels, and sizes are

shrunk and extended up to 20 percent. Therefore, by

dispersing particles in a relatively smaller region

instead of the whole space, searching procedure can

be definitely accelerated.

Then particles move around, communicate and

share information among the society, follow the

better directions of their neighbors, and converge to

ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics

104

the optima. This process is automatic and

independent on knowledge of image contents. After a

number of iterations, particles cluster around one or

several optimal points in the space, which correspond

to some regions with varied locations and sizes.

These regions are candidates for the Bayes filter,

which will be discussed in later section.

4.4 Fitness Function

The behaviors of particles are guided by the

associated fitness function, which defines the search

criteria underlying the PSO searching algorithm. In

terms of object tracking, fitness function can be

defined as a function of features of the tracked object.

Lots of features are used for objects detection and

tracking, including color, texture, shape and motion,

which can be employed independently or several

features can be combined together. In this paper, the

appearance histogram is applied to construct the

fitness function.

First, images are transformed from RGB format

into HSV format, and the later one is more natural

for people’s eyes. Then, the values of hue are

abstracted to build the histogram. Such histogram

refers to the gradation of color within the visible

spectrum. When a PSO-based searching algorithm is

applied, each particle at every moment is associated

with a histogram. The best matched one can be

obtained by comparing these histograms with the

target histogram. Therefore, a special criterion is

required to measure the similarity between the

searched window and the target window, which

means a method to measure the distance between two

histograms is required.

In statistics (T. Kailath, 1967), the Bhattacharyya

Coefficient measures the similarity of two discrete

probability distributions. It is a divergence-type

measure that can be seen as the scalar product of the

two vectors having components as the square root of

the probability of the points x ∈ X. It thereby lends

itself to a geometric interpretation: the Bhattacharyya

Coefficient is the cosine of the angle enclosed

between these two vectors. Therefore, the

Bhattacharyya Coefficient is used to measure the

similarity between these two histograms, which is

defined as:

∑

∈

gigi

xHxHHHBC )()(),( (4)

Where represents the histogram of particle i,

represents the histogram of the target, and X

denotes the distribution domain, which is the range

of hue values from 0 to 255. and are

pixel numbers with a specific hue value x for the

particle and target, respectively.

)(xH

By using (4), the distance between two

histograms can be defined as (D. Comaniciu, V.

Ramesh, and P. Meer, 2004):

),(1),(

gigi

HHBCHHD −= . (5)

This distance is invariant to the scale of the target,

while the popular used histogram intersection is scale

variant (M.J. Swain, D.H. Ballard, 1991). The

smaller this distance is, the better the particle is

matched with the target object. Thus given the target

histogram, the fitness function for particle i is

inversely proportional to the distance between

and :

),(/1),(

gii

HHDgpF =

(6)

The higher the fitness value, the more similar the

corresponding area is with the target.

4.5 Bayes-Based Filter

For each frame a motion vector V can be calculated

according to a motion trajectory of the tracking

window. The motion vector is zero in the first frame.

And for others, it is the shift from the previous

position to the current one.

Given the previous tracking window associated

with the target histogram and the motion

vector

{

}

VH ,

, where represents the motion

vector of target. The PSO-based searching algorithm

returns a set of candidate windows, which can be

represented by

{

}

miVH

,...,2,1|,

, where

represents histograms of particle i, represents

motion vectors of particle i, and m is the number of

the selected candidate windows. All of these

candidate windows are good enough in terms of

appearance features and their fitness values are

higher than a preset threshold.

According to Bayes law, the problem can be

described as:

),(

),(),|,(

),|,(

iiiigg

ggii

VHp

VHpVHVHp

VHVHp =

(7)

),|,(

ggii

VHVHp

represents the condition

probability of a particle with

{

given

}

VH ,

{

}

VH , .

represents the probability of the target

window, which is same for all particles.

represents the back projection

),(

VHp

),|,(

iigg

VHVHp

BAYES-BASED OBJECT TRACKING BOOSTED BY PARTICLE SWARM OPTIMIZATION

105

from candidates to the previous tracking window.

Since all of particles can point back to the target

window in different ways, it is hard to tell which

particle is the most possible one without any

predefined knowledge of the image environment. In

this paper, we simply assume that

all , i = 1,2,…,m, are equal.

However this assumption may not hold in some

practical applications, for instance a mobile vision

system where the previous motion trajectory of the

mobile platform would provide more information for

the back projection, which will be investigated in our

future work.

),|,(

iigg

VHVHp

Considering that the PSO-based searching

algorithm returns all of candidates which are good

enough in appearance histogram, it is reasonable to

ignore the histogram here and simplify (7) as:

)()|(

igi

VcpVVp = , (8)

where c is a positive constant factor, and

represents the probability of a particle on the

motion trajectory. According to the inertia of motion,

depends on the distance between and .

The closer two vectors are, the higher the possibility

of the corresponding particle, which makes (8) as the

following equation:

)(

),(

)()|(

igi

VVD

VcpVVp ==

(9)

where k is a positive factor. If two vectors are

shifted to the same original point, the distance

between two vectors turns into the distance between

two points, where Euclidean distance can be

calculated.

5 EXPERIMENTAL RESULTS

To evaluate the proposed algorithm, some video clips

from PETS database are applied in this paper. The

program is written in C++ using OPENCV library,

running on a Pentium4 desktop. Most data come

from a surveillance system with a stationary camera.

Figure 2 shows the process of identifying moving

objects by motion detection, where pictures from left

to right are true data, foregrounds, and backgrounds,

respectively. If there is no moving object, as shown

in Figure 2(a), the background is the same with true

image and the foreground is empty since no object is

detected. With some general preprocessing, the noise

can be depressed and the model of background can

be enhanced. When a car drives in, it is detected and

recognized as an object. As shown in Figure 2(b), a

car shape appears in the foreground while the

background keeps the same with the true image. For

most testing data with static background, motion

detection can detect moving objects quickly. For

those testing data under dynamic environment, some

pre-knowledge of objects, such as moving behaviors,

would help to improve the detection performance.

Figure 3 shows the procedure of the proposed

PSO algorithm searching for candidate windows. A

number of particles are distributed around the target

according to the tracking window of previous frame

in Figure 3(a). Due to the uncertainty of the object

movement, initially, these windows are set up as

different sizes and locations near the detected object

using motion detection. Then particles start to move

around and eventually converge to some optimal

points under PSO rules. Figure 3(b) shows these

optimal points, which are good candidates of tracking

windows. As shown in Figure 3(b), it is obviously

that these candidate windows are much closer to the

car compared with those initial windows in Figure

3(a), which demonstrates the efficiency of the PSO-

based searching algorithm. Then Bayers filter is

applied to select the best match from those good

candidates, as shown in Figure 3(c). Usually, the

PSO-based searching algorithm converges quickly.

In our experiments, initially 20 windows are

generated, then after 10 to 15 time steps, those

windows cluster to the object.

To evaluate the robustness of the proposed

tracking method under occlusion, another experiment

is carried out as shown in Figure 4. First, a white car

drives in and is detected as the target by a blue

rectangle window as shown in Figure 4(a). Then, the

white car traverses the scene and is occluded by a

block of texts in the image, as shown in Figure 4(b)

and (c). During the occlusion, the tracking window

changes with scenes, but still tracks the car. As

shown in Figure 4(b), when the car starts moving into

the block, the tracking has almost the same size with

the one in Figure 4(a). Under the influence of the

block, the tracking window shifts a little and shrinks.

But the object is still locked. When the car moves

away as shown in Figure 4(d), the window becomes

smaller until disappeared. It can be seen that the

tracker can still lock the object under occlusion.

The above experiments demonstrate the proposed

algorithm is efficient and robust. However under

some complex situations, such as dynamic

background, more robust motion detection is

required. For some noisy videos, the tracking

window may be lost due to frame skips. A recovery

algorithm may need to increase the system reliability.

ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics

106

6 CONCLUSION

In this paper, a robust adaptive-window based

tracking algorithm is proposed to automatically

detect and track moving objects. First, a motion

detection algorithm is applied to detect the moving

object. Then a PSO-based searching algorithm comes

to search for good candidates of adaptive tracking

windows with parameters on the new fame. Last,

Bayes-based filter is employed to identify the best-

matched tracking window under the motion

constraints of the tracked object. The experimental

results demonstrate that the proposed algorithm is

robust and efficient in some popular used video data

There are still several issues remained and need

to be improved and extended in the future work. The

first one is to investigate new object detection

approaches under dynamic environment where the

background of the image and illumination conditions

may be dramatically changed and the motion

detection and histogram-based method applied in this

paper will not be reliable any more. The second one

is to concrete the Bayes filter using some predefined

knowledge of the tracked targets.

(a) (b)

Figure 2: Motion detection to recognize objects, (a) to (b) from left to right.

(a) (b) (c)

Figure 3: Tracking procedure using PSO-based searching.

(a) (b) (c) (d)

Figure 4: Tracking under occlusion.

BAYES-BASED OBJECT TRACKING BOOSTED BY PARTICLE SWARM OPTIMIZATION

107

REFERENCES

G. Welch and G. Bishop, “An Introduction to Kalman

Filter”, SIGGRAPH 2001, Course 8, 2001.

M. Isard, “Visual Motion Analysis by Probabilistic

Propagation of Conditional Density”, D.Phil. Thesis,

Oxford University, 1998.

S. Maskell and N. Gordon, “A Tutorial on Particle Filters

for On-line Nonlinear/Non-Gaussian Bayes’Law

Tracking”, 2001.

D. Comaniciu, and P. Meer, “Mean shift: a robust approach

toward feature space analysis,” IEEE Trans. Pattern

Analysis Machine Intel., Vol. 24, No. 5, pp. 603-619,

2002.

J. Kennedy, R. C. Eberhart, and Y. Shi, Swarm intelligence,

San Francisco, Morgan Kaufmann Publishers, 2001.

T. Olson and F. Brill. “Moving Object Detection and Event

Recognition Algorithms for Smart Cameras”, in Proc.

DARPA Image Understanding Workshop, pp. 159-175,

1997.

Yuri Owechko, Swarup Medasani, and Narayan Srinivasa,

"Classifier Swarms for Human Detection in Infrared

Imagery,” in 2004 Conference on Computer Vision and

Pattern Recognition Workshop (CVPRW’04), Vol. 8,

pp. 121, 2004.

Reza Akbari, Mohammad Davarpanah Jazi, and Maziar

Palhang “A Hybrid Method for Robust Multiple

Objects Tracking in Cluttered Background”, Vol. 1,

24-28, ICTTA ’06.

Luis Anton-Canalis, Mario Hernandez-Tejera, and Elena

Sanchez-Nielsen etc. “Particle Swarms as Video

Sequence Inhabitants For Object Tracking in Computer

Vision”, Vol. 2, pp. 604-609, ISDA ’06.

R. Eberhart and J. Kennedy. "A New Optimizer Using

Particles Swarm Theory", in Proc. Sixth International

Symposium on Micro Machine and Human Science,

Nagoya, Japan., IEEE Service Center, Piscataway, NJ,

39-43, 1995.

J. Kennedy and R.C. Eberhart, "Particle Swarm

Optimization", in Proceedings of IEEE International

Conference on Neural Networks, 1942-1948, 1995.

Kennedy, J. and R. Eberhart, “A discrete binary version of

the particle swarm algorithm,” in Proceedings of the

Conference on Systems, Man and Cybernetics, pp.

4104-4109, Piscataway, New Jersey, IEEE Press, 1997.

J. R. Jain and A. K. Jain, “Displacement measurement and

its application in interframe image coding”, IEEE

Trans. Commun. Vol. COM-29, no.12, pp.1799-1808,

Dec. 1981.

T. Kailath, “The Divergence and Bhattacharyya Dis-tance

Measures in Signal Selection," IEEE Trans. Com-mun.

Tech., COM-15:52-60, 1967.

D. Comaniciu, V. Ramesh, and P. Meer, “Real-Time

Tracking of Non-Rigid Objects using Mean Shift”,

CVPR, Vol. 2, pp. 142-149, 2000.

M. J. Swain, D.H. Ballard, “Color Indexing," Intern.

J.Comp. Vis., 7(1):11-32, 1991.

A. Leykin and R. I. Hammoud, “Robust Multi-Pedestrian

Tracking in Thermal-Visible Surveillance Videos”,

IEEE Computer Vision and Pattern Recognition

Workshop, OTCBVS'06, New York,USA, 17-22 June

2006 pp:136 – 141

Y. Zhang, S. J. Kiselewich, W. A. Bauson, and R.

Hammoud, “Robust Moving Object Detection at

Distance in the Visible Spectrum and Beyond Using a

Moving Camera”, IEEE Computer Vision and Pattern

Recognition Workshop (CVPRW), on Object Tracking

and Classification Beyond the Visible Spectrum, New

York, USA, 17-22 June 2006, pp: 1331-137

ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics

108