APPLYING DYNAMIC MODEL FOR MULTIPLE MANOEUVRING TARGET TRACKING USING PARTICLE FILTERING

International Journal of Information Technology, Control and Automation (IJITCA) Vol.2, No.4, October 2012
DOI:10.5121/ijitca.2012.2404 37
APPLYING DYNAMIC MODEL FOR MULTIPLE
MANOEUVRING TARGET TRACKING USING
PARTICLE FILTERING
Mohammad Javad Parseh
1
and Saeid Pashazadeh
2
1
M.Sc. Student of Artificial Intelligence
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
m_paresh89@ms.tabrizu.ac.ir
2
Assistant Professor of Software Engineering
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
pashazadeh@tabrizu.ac.ir
ABSTRACT
In this paper, we applied a dynamic model for manoeuvring targets in SIR particle filter algorithm for
improving tracking accuracy of multiple manoeuvring targets. In our proposed approach, a color
distribution model is used to detect changes of target's model. Our proposed approach controls
deformation of target's model. If deformation of target's model is larger than a predetermined threshold,
then the model will be updated. Global Nearest Neighbor (GNN) algorithm is used as data association
algorithm. We named our proposed method as Deformation Detection Particle Filter (DDPF). DDPF
approach is compared with basic SIR-PF algorithm on real airshow videos. Comparisons results show
that, the basic SIR-PF algorithm is not able to track the manoeuvring targets when the rotation or scaling
is occurred in target's model. However, DDPF approach updates target's model when the rotation or
scaling is occurred. Thus, the proposed approach is able to track the manoeuvring targets more efficiently
and accurately.
KEYWORDS
Multiple manoeuvring target tracking, particle filtering, video tracking, dynamic models.
1. INTRODUCTION
Multi target tracking is one of the important research areas in computer vision. Human tracking in
monitoring applications and aircraft tracking in military applications are known as important
applications areas of multi target tracking. Multi target tracking includes two main stages: 1) pose
estimation and 2) data association. Kalman filtering and grid-based methods are the most
important methods for pose estimation that are used in linear systems with Gaussian noises. For
linear-Gaussian systems, Kalman filter is an analytical and optimal method. Particle filtering is a
numerical and suboptimal method that is used for pose estimation. Unlike the Kalman filtering,
Particle filtering method can be used in non-linear system with non-Gaussian noises. This is the

38
most important advantage of particle filtering over other methods. Estimation of manoeuvring
target's state is one of the most important problems in pose estimation. Multiple Model (MM)
approach is a common solution to resolve this problem [1-3]. Target's posteriori density function
is represented by weighted sum of the output of several parallel filters in MM approach [4]. A
real-time approach is presented in [15] for tracking a manoeuvring target by incorporating
dynamic components in the target model. They employed particle filter based tracker. This
tracker exploits a first order dynamic model and continuously performs adaptation of model noise
for balancing uncertainty between the static and dynamic components of the state vector. The
presented experimental evaluation shows that this approach is particularly effective in video
sequences where appearance and motion have quick changes and there are partial or full target
occlusions.
Data association algorithm associates each measurement to track. In cluttered environments,
measurements may arise from detected targets or clutter. From the view of implementation,
Global Nearest Neighbor (GNN) is the simplest method for data association [5]. Usually, this
method is more efficient in clutter-free environments. One of the most common data association
methods is Joint Probabilistic Data Association (JPDA) [6]. In this method, the probability of
each measurement associated with each track is calculated. Multiple Hypothesis Tracking (MHT)
is the most complex method for data association that is efficient for cluttered environments [7]. In
MHT method, each hypothesis is a measurement-to-track association. Set of all hypothesizes is
considered, and likelihood of each hypothesis is calculated. One of the best methods for obtaining
best hypothesis is presented by Nummiaro et. al. [8]. In MHT method, the best hypothesis is
chosen through the set of all hypothesizes by using a statistical approach, whereas in the JPDA
method, an average of target state is calculated on all hypothesizes. This is the main difference
between JPDA method and MHT method. In this paper, DDPF algorithm is proposed by
improving the basic SIR algorithm for tracking the multiple manoeuvring targets using a dynamic
model for targets.
Remaining sections of this paper is organised as follows: In section 2, particle filter method is
introduced briefly. In section 3, basic SIR particle filter algorithm is introduced and in section 4,
target representation model is explained. In section 5, a dynamic equation for manoeuvring target
and in section 6, likelihood function is presented. In section 7, we briefly describe data
association algorithm. In section 8, a new method is presented for deformation detection. In
section 9, conditions of experiments is presented. In section 10, we compare results of our
proposed DDPF method and basic SIR algorithm in multiple manoeuvring target tracking.
Finally, in section 11, we’ll discuss about and make conclusion from experimental results.
2. PARTICLE FILTER
Appropriate pose estimation method should be chosen according to tracking conditions. If the
system is linear and system noise is Gaussian, Kalman filter is most appropriate method to pose

39
estimation. But, if the system's dynamic equation is non-linear and system noise is non-Gaussian,
an approximation technique should be used. One of the most common numerical and
approximation methods to pose estimation is sequential Monte Carlo [9]. This technique is known
as particle filter [10] or Monte Carlo [11]. Let assume Xt denotes the target's state at time t, and Zt
denotes the set of all measurements zt up to time t. The main idea in particle filtering is
representation of probability distribution function (pdf) by a set of weighted samples. Suppose
that ( ){ }( ) ( )
, | 1,...,n n
S s n Nω= = is a set of weighted samples where ( )
1
1
N
n
n
ω
=
=∑ , then the posteriori
density function is represented as follows.
( ) ( )
1
|
N
i i
t t t t t
i
p X Z X Xω δ
=
≈ −∑ (1)
where i
tω is the weight of sample i at time t which is calculated by following equation.
( ) ( )
( )
1
1
1
| |
| ,
i i i
t t t ti i
t t i i
t t t
p z X p X X
q X X z
ω ω
−
−
−
∝
(2)
where ( ).q is importance density function which is used in sampling step. This method is called
Sampling Importance Sampling (SIS). In the equation (1), δ is Kronecker delta function which
is defined as follows:
( )
1 0
0 0
t
t
t
δ
=
= 
≠ (3)
In this paper, target’s state is defined as follows:
( ){ }, | 1,...,k k
t t tX x y k T= =
(4)
where k
tx and k
ty are coordinates of target k at time t in video screen, and T is the number of
targets. The particle filter method includes two main stages: prediction and update. In the
prediction stage, particles of next step are obtained by using current particles and state transition
equation. In the update stage, a weight is assigned to each particle using the available
measurements. After the weight assignment, posteriori density function is calculated by equation
(1). Then the target's next state is computed by a weighted average on the next step particles using
following equation:
1
ˆ
N
i i
t t t
i
X Xω
=
= ∑ (5)

40
3. SIR PARTICLE FILTER
One of the big challenges in implementation of SIS particle filter algorithm is appropriate choice
of importance density function ( ).q . This function is used in sampling step. Accuracy of pose
estimation increases by increasing the similarity between importance density function and
posteriori density function. Another problem in implementation of particle filter is degeneracy
problem. After several iteration of particle filter algorithm, some of the particles have very small
weight. This means that, some of the particles don’t have effect on calculation of posteriori
density function. This is an adverse event. There are two solutions to overcome the degeneracy
problem: 1) appropriate choice of importance density function 2) resampling. Both of these
solutions are used in SIR approach. State transition function is considered as importance density
function.
( ) ( ) ( )1 1| , | ,t
i i
t t t t t X t tq X X z p X X X− −= = Ν ∑
(6)
This is a suboptimal choice for importance density function. After state estimation in each step of
SIR algorithm, current samples are resampled. After the resampling step, weights of samples are
equal to 1
N
(N is the number of samples). If the state transition function is considered as
importance density function, then the particle weights are calculated by following equation:
( )1 |i i i
t t t tp z Xω ω −∝
(7)
Equation (7) should be modified as follows, because weights of samples are equal to 1
N
after
the resampling step.
( )|i i
t t tp z Xω =
(8)
where ( )| i
t tp z X is likelihood of each particle at time t.
4. TARGET REPRESENTATION MODEL
In this paper, each target’s state is represented by a rectangle. Center of the rectangle denotes the
target’s current state, and dimensions of the rectangle are adjusted by target's size. Fig.1 shows
the representation model of targets. In this figure, xH and yH are dimensions of surrounding
rectangle of target object, and ( ),x y denotes the current state of each target. Model of each target
consists of the target's state and dimensions of surrounding rectangle. If changes of the model
become larger than the predetermined threshold then target’s model will be updated. When the
target model is updated, both target state and dimensions of corresponding rectangle will be
updated.

41
Figure 1. Target representation model.
5. SYSTEM DYNAMIC EQUATION
In this paper, we used real-world videos. These videos are extracted from real airshow videos that
obtained from internet. In these videos, camera is movable. Sudden motion of camera causes
undesirable motion in trajectory lines of targets. So, a discrete-time random walk process is
considered as state transition equation for each target. In random walk model, the state transition
equation is defined as follows.
1t t tX X υ−= +
(9)
where ( )0,t pυ Ν Σ is a two dimensional Gaussian noise. Probability of state transition of each
target is defined as following equation.
( ) ( )1| ,tt t X t tp X X X− = Ν Σ
(10)
6. LIKELIHOOD FUNCTION
We used combination of two functions to define the likelihood function in this study. First
function calculates the likelihood according to difference between corresponding pixels in
grayscale mode [16]. Second function calculates the likelihood of corresponding regions using a
color histogram in RGB mode [13]. The likelihood function is defined as following equation:
( ) ( )
2
int| i
t t ensity colorp z X p p= +
(11)
In section 6.1, grayscale module is briefly explained, and an equation is defined to calculate
intensityp . In section 6.2, RGB module is introduced and an equation is defined to compute colorp .
6.1. Grayscale Module
Difference in intensity of corresponding pixels is used to calculate the similarity between target’s

42
region and each particle's region. If S is the set of pixels inside the target’s region and jC is the
set of pixels inside the jth
particle’s region, then we calculate similarity of these regions using
following equation [18]:
( )
1
int
I
i i
j
i
x y
T C
H H
ensityp e
=
 
 −
 
− 
× 
 
 
∑
=
(12)
where xH and yH are the dimensions of rectangular region of target object, and I is the number
of pixels inside this region.
6.1. RGB Module
A color distribution model is used as measurement model in some studies [14]. This model is an
8×8×8 histogram in RGB color space. This histogram is calculated for target's rectangular region.
Distance of each pixel from center of region affects values of histogram. In [14], the interval
(0,255) is divided to 8 equal parts, and an 8×8×8 histogram is used. For reducing computational
complexity, we used a 4×4×4 histogram. In this color distribution model, efficacy of each pixel
on values of histogram reduces as it is placed far from the center of region. Epanechnikov kernel
function is used to indicate this property in the histogram, which is defined as follows:
( )
2
1 1
0
r r
k r
otherwise
 − <
= 
 (13)
where r is the distance of each pixel from the center of region. The color distribution model is
represented by a histogram which is shown in following equation:
{ }( )
1,...,
u
s s u m
p p
=
=
(14)
where s denotes the location of an arbitrary pixel, and m is the number of histogram bins. Number
of histogram bins is considered as m=64. The histogram value in location s is obtained by
following equation:
( )( )( )
1
I
iu
s i
i
s X
p f k h X u
a
δ
=
 − 
= − 
 
∑ (15)
where I is the number of pixels inside the target’s region, and . is considered as Euclidean
distance. 2 2
x ya H H= + , where xH and yH are the dimensions surrounding rectangle of target
object. In equation (15), f is a normalization coefficient which is defined as follows:

43
1
1
I
i
i
f
s X
k
a=
=
 − 
 
 
∑ (16)
Bhattacharyya distance is used to calculate the difference of two histograms using following
equation:
[ ]1 ,d p qρ= −
(17)
where [ ],p qρ denotes the similarity of p and q and is defined as follows:
[ ] ( ) ( )
1
,
m
u u
u
p q p qρ
=
= ×∑ (18)
Finally, the distance between, histogram of target’s region and histogram of each particle’s region
are used to assign the weight to each particle. colorp is calculated by following equation:
2
d
colorp e λ−
= (19)
where λ is a parameter that experimentally considered 25 in this paper.
7. DATA ASSOCIATION
Data association algorithm associates the available measurements to detected targets. In cluttered
environments, some of the obtained measurements may arise from clutter. A data association
algorithm must detect the measurements that are arisen from clutter. Some of the important data
association algorithms are Global Nearest Neighbor (GNN), Joint Probabilistic Data Association
(JPDA) and Multi Hypothesis Tracking (MHT). GNN is more efficient algorithm for clutter-free
environments. Our selected videos in this study had static background and our tracking
environment was clutter-free. We used GNN as data association algorithm. Each measurement is
associated to nearest track in GNN algorithm. Optimal association minimizes the sum of
distances between each measurement and it's corresponding track. The optimal association can be
obtained using a cost matrix and an optimization algorithm. Distance between each measurement
and track represents their cost in this matrix. We used Munkres optimization algorithm in our
study that was presented in [14].
This algorithm uses the cost matrix for determining the optimal association. We used three types
of measurements: 1) grayscale module, 2) RGB module and 3) target detection for deformation
detection. Grayscale module and RGB module don't require data association, because these are
calculated for each particle separately. Data association is only used for target detection.

44
In our application, number of targets is constant and predetermined. If the number of detected
targets in tracking environment is less than the predetermined number, then at least one of the
targets has been occluded by another one. In such case, target detection step is ignored during the
occlusion phenomenon. The color histogram and intensity of corresponding pixels are considered
as observation. Data association won't be applied during the occlusion phenomenon, because data
association is only used for target detection observation.
8. DEFORMATION DETECTION
Manoeuvring target tracking is one of the big challenges in target tracking. Target’s model
changes in occurrence of rotation or scaling. In such case, the target’s old model is not efficient
for weight assignment. Using static model for tracking manoeuvring targets don’t yield good
results. The MM approach is a commonly used approach to overcome this problem. In this paper,
a dynamic model is used to improve the results of SIR algorithm in multiple manoeuvring target
tracking. If changes of target’s model are larger than the predetermined threshold, then the model
should be updated. Old model of target is replaced by new one. Deformation detection is done
using a color histogram that is described in section 6.2. After several frames, the considered
targets are detected again by using a background subtraction algorithm. The color histogram is
calculated for region of each detected target as described in section 6.2. Suppose that, the newq is
the color histogram of new model of target and oldq is the color histogram of old model of target.
We used Bhattacharyya distance to detect the deformation of target model which is defined as
following equation:
[ ]1 ,new oldd q qρ= −
(20)
where [ ],new oldq qρ denotes the similarity of two histograms and obtains using following equation:
[ ] ( ) ( )
1
,
m
u u
new old new old
u
q q q qρ
=
= ×∑ (21)
when d T> , then changes of target’s model is larger than a predetermined threshold and the its
previous model should be replaced by new one. T is a threshold which is experimentally
considered as 0.12. Our proposed DDPF algorithm is described in Table 1.
9. CONDITIONS OF EXPERIMENTS
Our proposed DDPF algorithm and basic SIR algorithm are implemented by MATLAB 2011(b).
These algorithms are tested on a system with configuration: CORE i7 CPU, 4GB RAM and 1GB
Geforce VGA. The real-world videos are selected to test our proposed algorithm. These videos

45
are extracted from parts of real airshow videos. Format of these videos is "mp4", and their
dimensions are 240×320. It is assumed that the number of targets is constant and predetermined
in these videos. A background subtraction algorithm is used to detect the targets, because the
selected videos have partly static background. Background of selected videos should be reclusion.
Our proposed DDPF method is appropriate for tracking in clutter-free environment.
10. COMPARISON
In this section, we compare results of our proposed DDPF algorithm with basic SIR particle filter
algorithm in multi manoeuvring target tracking. Fig.2 shows that the basic SIR-PF algorithm is
not able to track the manoeuvring targets without using dynamic model for targets. Fig.3 shows
that DDPF algorithm efficiently tracks the manoeuvring targets. Comparison results show that the
basic SIR-PF algorithm misses the targets when the rotation or scaling occurs in target’s model,
whereas proposed DDPF method resolves this problem by updating the model when the rotation
or scaling occurs. Results of our proposed method are very sensitive to parameters settings. By
increasing the dimensions of videos, the computational complexity of our approach increases
drastically. We tested our proposed approach on other videos and their results are shown in Fig.4
and Fig.5.

46
Table 1. Deformation Detection Particle Filter algorithm
Figure 2. Result of basic SIR algorithm in multiple manoeuvring targets tracking without using
dynamic target's model.

47
Figure 3. Result of DDPF method in multiple manoeuvring targets tracking using dynamic
target's model.
Figure 4. Result of DDPF method in multiple manoeuvring targets tracking without occlusion.
Figure 5. Result of DDPF method in tracking two manoeuvring aircrafts with partial occlusion.

48
11. CONCLUSION AND DISCUSSION
In this paper, we proposed DDPF method by applying a dynamic model for targets in the SIR
based particle filtering algorithm for tracking multiple manoeuvring targets. Results illustrate that
the basic SIR-PF algorithm is not able to efficiently tracks the multiple manoeuvring targets when
the rotation or scaling is occurred in target model, whereas DDPF method will be able to track
multiple manoeuvring targets more robustly. Our studies showed that the results are very
sensitive to parameters settings. Extending this approach by adaptive parameters is under study as
our future work. One of the problems in multi manoeuvring target tracking is occlusion problem.
Resolving the occlusion problem for improving the results of proposed approach is our other
future work.
Figure 6. Result of DDPF method in tracking three manoeuvring aircrafts with partial occlusion.
Figure 7. Result of DDPF method in tracking two manoeuvring aircrafts with complete
occlusion.

49
REFERENCES
[1] H. A. P. Blom and Y. Bar-Shalom (1988) "The Interacting Multiple Model Algorithm for Systems
With Markovian Switching Coefficients", IEEE Trans. Automat. Control, Vol. 33, No. 8, pp. 780–
783.
[2] S. McGinnity and G.W. Irwin, (2000) "Multiple Model Bootstrap Filter for Manoeuvring Target
Tracking", IEEE Trans. Aerospace Electron. Syst., Vol. 36, No. 3, pp. 1006–1012.
[3] A. Doucet, N. J. Gordon and V. Krishnamurthy, (2001) "Particle Filters for State Estimation of Jump
Markov Linear Systems", IEEE Trans. Signal Process, Vol. 49, No. 3, pp. 613–624.
[4] Y. Ulker and B. Gunsel, (2012) "Multiple Model Target Tracking with Variable Rate Particle Filters",
Digital Signal Processing, Vol. 22, No. 3, pp. 417–429.
[5] Y. Bar-Shalom and T. Fortmann, (1988) "Tracking and Data Association", Academic Press.
[6] Y. Bar-Shalom, T. Fortmann, and M. Scheffe, (1980) "Joint Probabilistic Data Association for
Multiple Targets in Clutter", in Proc. of Conf. on Information Sciences and Systems, Princeton Univ.
[7] D. Reid, (1979) "An Algorithm for Tracking Multiple Targets", IEEE Trans. Automat. Control, Vol.
24, No. 6, pp. 84–90.
[8] I. Cox and S. Hingorani, (1996) "An Efficient Implementation of Reid’s MHT Algorithm and Its
Evaluation for The Purpose of Visual Tracking", IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 18, No. 2, pp.138–150.
[9] A. Doucet, N. de-Freitas and N. Gordon, (2001) "Sequential Monte Carlo Methods in Practice",
Springer-Verlag, New York.
[10] S. Arulampalam, S. R. Maskell, N. J. Gordon and T. Clapp, (2001) "A Tutorial on Particle Filters for
On-line Nonlinear/Non-Gaussian Bayesian Tracking", IEEE Trans. on Signal Processing, Vol. 50,
No. 2, pp. 174-189.
[11] M. Isard and A. Blake, (1998) "Condensation-Conditional Density Propagation for Visual Tracking",
International Journal on Computer Vision, Vol. 29, No. 1, pp. 5-28.
[12] B. Ristic, S. Arulampalam and N. Gordon, (2004) "Beyond the Kalman Filter", Artech House.
[13] K. Nummiaro, E. Koller-Meier and L. V. Gool, (2003) "An Adaptive Color-Based Particle Filter",
Image and Vision Computing (IVC), Vol. 21, No. 1, pp. 99–110.
[14] F. Bourgeois and J. C. Lassalle, (1971) "An Extension of The Munkres Algorithm for The
Assignment Problem to Rectangular Matrices", Communications of the ACM (CACM), Vol. 14, No.
12, pp. 802-804.
[15] D. Bimbo and F. Dini, (2011) "Particle Filter-Based Visual Tracking With a First Order Dynamic
Model and Uncertainty Adaptation", Computer Vision and Image Understanding, Vol. 115, No. 6, pp.
771–786.
[16] S. Fazli, R. Afrouzian and H. Seyedarabi, (2010) "Fiducial Facial Points Tracking Using Particle
Filter and Geometric Features", in Proc. of IEEE International Congress on Ultra-Modern
Telecommunications and Control Systems (IEEE ICUMT 2010), Moscow, Russia, pp. 396-400.
Authors
Mohammad Javad Parseh is M.Sc. student of Artificial Intelligence in University of Tabriz
in Iran. He received his B.Sc. in Software Engineering from Chamran university of Ahwaz in
2008. His research interests are Bayesian estimation, target tracking and state estimation,
ontology matching in semantic web, fuzzy decision making and operating system.
Saeid Pashazadeh is Assistant Professor of Software Engineering and chair of Information
Technology Department at Faculty of Electrical and Computer Engineering in University of
Tabriz in Iran. He received his B.Sc. in Computer Engineering from Sharif Technical
University of Iran in 1995. He obtained M.Sc. and Ph.D. in Computer Engineering from Iran
University of Science and Technology in 1998 and 2010 respectively. He was Lecturer in
Faculty of Electrical Engineering in Sahand University of Technology in Iran from 1999
until 2004. His main interests are development, modelling and formal verification of distributed systems,
computer security, particle filtering and wireless sensor/actor networks. He is member of IEEE and senior
member of IACSIT and Member of editorial board of journal of electrical engineering of University of
Tabriz in Iran.

APPLYING DYNAMIC MODEL FOR MULTIPLE MANOEUVRING TARGET TRACKING USING PARTICLE FILTERING

Recommended

More Related Content

What's hot (18)

Similar to APPLYING DYNAMIC MODEL FOR MULTIPLE MANOEUVRING TARGET TRACKING USING PARTICLE FILTERING (20)

More from IJITCA Journal (20)

Recently uploaded (20)

APPLYING DYNAMIC MODEL FOR MULTIPLE MANOEUVRING TARGET TRACKING USING PARTICLE FILTERING