A New Algorithm for Tracking Objects in Videos of Cluttered Scenes

International Journal of Information Technology, Modeling and Computing (IJITMC) Vol.1, No.2, May 2013
DOI : 10.5121/ijitmc.2013.1207 71
A New Algorithm for Tracking Objects in Videos
of Cluttered Scenes
Andres Alarcon Ramirez1
and Mohamed Chouikha2
1
Department of Computer Engineering, Howard University, Washington, DC, USA
alarconramirezandr@bison.howard.edu
2
Department of Computer Engineering, Howard University, Washington, DC, USA
mchouikha@howard.edu
ABSTRACT
The work presented in this paper describes a novel algorithm for automatic video object tracking based on
a process of subtraction of successive frames, where the prediction of the direction of movement of the
object being tracked is carried out by analyzing the changing areas generated as result of the object’s
motion, specifically in regions of interest defined inside the object being tracked in both the current and the
next frame. Simultaneously, it is initiated a minimization process which seeks to determine the location of
the object being tracked in the next frame using a function which measures the grade of dissimilarity
between the region of interest defined inside the object being tracked in the current frame and a moving
region in a next frame. This moving region is displaced in the direction of the object’s motion predicted on
the process of subtraction of successive frames. Finally, the location of the moving region of interest in the
next frame that minimizes the proposed function of dissimilarity corresponds to the predicted location of
the object being tracked in the next frame. On the other hand, it is also designed a testing platform which is
used to create virtual scenarios that allow us to assess the performance of the proposed algorithm. These
virtual scenarios are exposed to heavily cluttered conditions where areas which surround the object being
tracked present a high variability. The results obtained with the proposed algorithm show that the tracking
process was successfully carried out in a set of virtual scenarios under different challenging conditions.
KEYWORDS
Video object tracking, region of interest, cluttered conditions
1. INTRODUCTION
Video object tracking can be defined as the detection of an object in the image plane as it moves
around the scene. This topic has a growing interest for both civilian and military applications,
such as automated surveillance, video indexing, human-computer interaction (gesture
recognition), meteorology, and traffic management system [1][2][3].
There are two basic problems that a tracking system must resolve: the motion estimation and the
matching estimation. The motion estimation predicts the location of the most likely region in the
next video frame where the object being tracked may be placed. Commonly, this information is
not available; therefore, a mechanism to determine the fixed-size region surrounding the object
being tracked is needed. A technique widely used to resolve this problem in video tracking is
Kalman Filter (KF), which is an optimal recursive estimator of the state of a dynamic system. On
the other hand, matching estimation seeks to identify an object, which is being tracked in the
current video frame, inside a closed region in the next video frame. The closed region is predicted

72
by the motion estimation stage and corresponds to the zone where there is the highest probability
of finding the object of interest in the next video frame. The exact location of the object of
interest in the next frame is determined in the matching estimation stage by using information
extracted from the object itself in previous scenes. Matching estimation algorithms incorporates a
feature detection stage that constitutes the first step in different processing operations such as
image classification and segmentation. Feature detection is used by object tracking algorithms to
carry out a matching of the pixels from the object being tracked between two consecutive video
frames, and then determine the exact location of the object in the next frame.
The algorithm presented in this paper is an automatic video object tracking algorithm that uses a
region of interest (ROI) defined completely inside the object being tracked to carry out the
tracking process. To do this, the proposed algorithm guarantees that when the object to be
tracked suffers displacements, the ROI rebounds against the inner walls of the object being
tracked and stays inside this object. On the other hand, the paper also presents the designing of a
software platform that is used to create virtual scenarios where objects with different shapes and
sizes wander into scenes through the time. These virtual scenarios are useful to test the
performance of the proposed object tracking algorithm under different challenging problems such
as cluttered conditions and random movement.
This paper is organized as follows: In the Section 2, we describe the related work, in the Section
3, we present the novel proposed algorithm to track an object through a video. The section 4, on
the other hand, describes the software platform used o create virtual scenarios to test video object
tracking algorithms. In the Section 5, it is shown the obtained results. Finally, the Section 6
presents the conclusions of this work.
2. RELATED WORK
The process of tracking an object in a sequence of frames is directly dependent on the object’s
representation being used. Some representations, for example, use interest points to identify the
object to be tracked [4]. These interest points can be detected by using information based on
differentiation operators [5][6], where changes in intensity between two adjacent pixels can
emphasize the boundaries of the object of interest in the image [6]. Other object’s representations
use its silhouette or contour to extract information about the general shape of the object [7] [8].
Once the contour or the inner region of a given object is identified, different characteristics are
then extracted and used as features to be tracked between two consecutive frames.
Cross-correlation, on the other hand, was used in [9] to implement a face tracking algorithm for
video conferencing environment. This method compares a region of the image with a known
signal extracted from the object of interest, and then a measure of similarity, which allows to
determine the exact position of the object being tracked in the next frame, is obtained between the
two signals. In [10] is presented a methodology for video object tracking that is constituted by
four steps, namely, background subtraction, candidate object identification, target object
selection, and motion interpolation. Firstly, the background, which is available, is subtracted to
the current frame to identify the object being tracked; additionally, the background is updated
from time to time whenever there is a permanent change in it. Then, a threshold is applied over
the image, which is the result of subtracting the current frame from the background, to generate a
new binary image with the candidate objects from which the target object is selected using
histogram matching. Finally, motion interpolation determines the displacement of the object from
one frame to another.
Mean-shift (MS) is another technique of video object tracking that is based on primitive
geometric shapes [11]. At the beginning, it is defined a region of interest around the object to be

73
tracked in the current frame, and then it is started an iterative process based on comparing the
histogram of the region of interest in the current frame with the histograms obtained from
candidate regions in the next frame where there exist the chances of finding the object being
tracked. Finally, the location of the object being tracked in the next frame is defined by the
candidate region whose histogram presents the greatest similarity with the histogram from the
region of interest which surrounds the object in the current frame.
Scale-Invariant Feature Transform (SIFT), which is used for finding local points [12], was
integrated to MS to create a methodology that jointly employs point feature correspondences and
object appearance similarity [13]. Chakraborty and Patra presented a Kernel-based algorithm that
uses segmentation techniques to determine the target localization [14]. Babu et al. [15], instead of
using a single mean shift tracker, used multiple mean shift tracking points.
Chun-Te et al. [16] used projected gradient to help multiple inter-related kernels in finding the
best match during tracking under predefined constraints. On the other hand, multiple kernels were
incorporated into a Kalman filtering-based tracking system [17]. In their design, not only the state
transition matrix but also the noise covariance matrix used in Kalman filtering is dynamically
updated. Liu et al. [18] proposed a pixel classification approach based on Markov random field
MRF to track objects in video sequences, where kernel density estimation founded on
nonparametric models was used to represent both video objects and background. Additionally,
spatial context and temporal coherency modeled in MRF are exploited to ensure a more robust
segmentation performance.
Hossein and Bajie proposed a framework [19] for tracking moving objects based on spatio-
temporal Markov Random field, and where are taken into account the spatial and temporal
aspects of the object’s motion. Amer presented an automatic object tracking algorithm based on
the matching of features in successive frames [20]. Initially, the objects being tracked are
segmented and their spatial and temporal features are computed. Then, using a nonlinear two-
stage voting strategy, each object of the previous frame is matched with an object of the current
frame creating a unique correspondence.
However, the previous tracking algorithms have shown to have difficulties when tracking objects
in videos of cluttered scenes. Namely, when the pixels from the region around the object being
tracked present intensity variations across the video.
3. TRACKING APPROACH
The proposed algorithm is constituted by two important stages: the stage of estimation of the
direction of the object’s motion and the matching process. The first stage analyzes the changing
areas generated as result of the object’s motion, specifically in regions of interest defined inside
the object being tracked in both the current and the next frame. The latter stage determines the
location of the object being tracked in the next frame by evaluating a function of dissimilarity that
is minimized using information extracted from the object being tracked in the current video
frame.
3.1. Estimation of the Direction of the Object’s Motion
The proposed new strategy seeks to simplify the problem of tracking. To do this, the direction of
the object’s motion is estimated by employing the changing areas generated from the object’s
motion itself between the current frame and the second frame, and where only the pixels
belonging to two regions of interest defined in the first and the second frame respectively are
taken into account. The region of interest in the first frame is defined in such a way that this

74
region is completely inside the object of interest, whereas the second region of interest is defined
in the same coordinates than the first one but in the second frame.
Figure 1: Motion of a background object and the object to be tracked between two consecutive images, (a)
Current frame of the video, (b) next frame of the video, (c) region of interest located in the first frame, (d)
region of interest located in the second frame, (e) resulting image of subtracting two consecutive images, (f)
region of interest located in the resulting image of subtracting two consecutive image.
The Fig. 1 (a) and the Fig 1 (b) show a simple scenario where both a background object and the
object being tracked are displaced between the first and the second frame. Additionally, it is
shown the location of the region of interest in the first frame and the second frame respectively.
The Fig. 1 (c) shows that the environment around the object of interest in the first frame is
ignored. Similarly, the Fig. 1 (d) shows the zone in the second frame where the object of interest,
the background object, and the background are delimited by the region of interest.
The Fig. 1 (e), on the other hand, shows the resulting image of subtracting the first frame shown
in the Fig. 1 (a) of the second frame shown in the Fig. 1 (b). This resulting image presents two
regions of pixels of non-null values on a background of pixels of null values; the regions with
non-null pixels are generated as result of the movement of both the object of interest and the
background object. Nevertheless, these regions of pixels as a whole offer no clue about the nature
of the movement carried out by the object of interest.
On the other hand, if we define the same region of interest used in the first two frames but in the
image shown in the Fig. 1 (e), we will obtain a new image which reduces considerably the
complexity of the image shown in the Fig. 1 (e). This new image, which is shown in the Fig. 1 (f),
only takes into account the behaviour of the object being tracked in the region of interest, and
ignores the environment that surrounds the region of interest. To determine the direction of the
object’s motion, we select the region of interest defined in the object being tracked, and then it is
constituted two sets of pixels such as follows:
(a) (b) (c)
(d) (e) (f)

75
( ) ( ){ }0,),(|,1 ≠∈= yxFandRyxyxS (1)
Thus, the set, S1, represents the group of coordinates of the pixels in the region of interest, R,
where the image, F(x,y), has non-null values, i.e., F(x,y)≠0. Note that, F(.), represents the
resulting image of subtracting two consecutive images. The second set is constituted by the
coordinates of all pixels which are in the region of interest such as follows:
( ){ }RyxyxS ∈= ),(|,2 (2)
The groups of pixels which constitute the sets, S1 and S2, are shown in the Fig. 2 (a) and the Fig. 2
(b) respectively.
Figure 2: Analysis of the region of interest in the resulting image of subtracting two consecutive frames,
(a) region of pixels used to calculate the centroid Pn, (b) region of pixels used to calculate the centroid Pm,
(c) locations of the centroids Pn and Pm; vector that defines the direction of object’s motion.
On the other hand, if we use the Equation (3) to calculate the average of the coordinates that
constitute the group, S1, which was defined by the Equation (1), we will obtain the point, Pn,
which represents the centroid of the group,S1.
Figure 3: Vectors obtained under different directions of motion of the object being tracked.
In the same way, the centroid of the group, S2, which is represented by the point, Pm, is calculated
using the Equation (4). The locations of the centroids, Pn and Pm, in the region of interest are
shown in the Fig. 2 (c).
(a) (b) (c)

76
(3)
(4)
The two points, which correspond to the centroids of the sets, S1 and S2, constitute a vector whose
direction determines the orientation of the object’s motion. In other words, the vector which
connects the centroid, Pn, to the centroid, Pm, has the same direction than the object being
tracked. This vector is defined by the following equation.
(5)
Finally, the angle of the vector which determines the direction of the object’s motion is calculated
using the Equation (6).
(6)
The Fig. 3 presents the regions of interest in the images generated from subtracting two
consecutive frames, and where the object being tracked is moved in different orientations. Each
of these regions shows the location of the centroids, Pn and Pm, which were calculated from the
sets, S1 and S2 respectively. In the same way, the two centroids constitute the vector which
determines the orientation of movement of the object being tracked.
Once the direction of the object being tracked has been determined, it is started an iterative
process which seeks to determine the location of the object being tracked in the next frame
(second frame). To do this, it is used the region of interest which was defined in the current frame
(first frame), R1, to determine the direction of the object’s motion. This region of interest is
located totally inside the object being track such as it is shown in the Fig. 4 (a). At the same time,
it is defined in the second frame a second region of interest, R*, with the same shape, size, and
location of the first region of interest used in the current frame. The initial location of the region
of interest, R*, in the second frame is shown in the Fig. 4 (b).
Figure 4: Vectors obtained under different directions of motion of the object being tracked.
During each iteration of the iterative process, the region of interest, R*, is displaced across the
second frame in the direction of the object’s motion which was initially estimated. The Equations
(7) and the Equation (8) describe the movement of the region of interest, R*.
(7)
(8)
nm PPV −=
IVangle )(=
( ) ( )
( )
∑∈
==
N
Syx
iinnn
ii
yx
N
yxP
1,
,*
1
,
( ) ( )
( )
∑∈
==
N
Syx
iinnm
ii
yx
N
yxP
2,
,*
1
,
(a) (b) (c) (d)
)cos(* kx=∆
)sin(* ky =∆

77
where, k is an integer which takes the values of 0,1,2,.., S. The parameter, S, is a constant which
represents the maximum possible displacement of the object being tracked; its value is defined by
the user and depends on the nature of the video being processed. Finally, the parameter, θ, is the
angle of the vector which defines the direction of the object’s motion. This angle is calculated
using the Equation (6).
On the other hand, at each iteration of the iterative process which seeks to establish the
displacement of the object being tracked, the region of interest, R*, located in the second frame is
compared with the region of interest, R1, located in the first frame. The Equation (9) presents the
function of dissimilarity, M(∆x, ∆y), which is used to compare these two regions.
(9)
where the parameter, L, represents the number of pixels which constitutes the region of interest ,
R1 . The function, U(.), is defined by the Equation (10).
(10)
It is important to mention that that the function, M(.), depends on the parameters, ∆x and ∆y,
which correspond to the horizontal and vertical displacements of the second region of interest, R*.
In general, the function, M(.),compares one-to-one the pixels located in the first region, R1 , with
the pixels located in the second region, R*, to determine the number of different pixels between
these two regions. Thus, to the end of the iteration process, we will obtain a set of values for the
function, M(.), for different values of displacement. The Fig. 4 (b) to (f) present the graphical
representation of the movement of the region of interest, R*, across the second frame following
the direction of the object’s motion.
To the end of the iterative process, it is selected the pair of values for ∆x and ∆y that after being
evaluated in the function, M(.), obtains the minimum value for this function among a total of ‘S’
possible values. The pair of values selected defines the displacement of the region of interest, R*,
which corresponds to the same displacement carried out by the object being tracked between the
current and next frame. The iterative procedure described above constitutes a process of
minimization that can be described by the following expression:
(11)
At the end of the process of minimization, the region of interest in the current frame is updated to
the new location defined by the pair of values, ∆x and ∆y, that minimizes the function, M(.), such
as follows:
(12)
(13)
( )
( )
( )∑=












∆+∆+
−
=∆∆
L
i ii
ii
yyxxR
yxR
UyxM
1
1
,*
,
,
( )
( )
( ) 



















∆+∆+
−
=∆∆ ∑=
L
i ii
ii
yyxxR
yxR
UyxMMin
1
1
,*
,
,
xXX ii ∆+=
yYY ii ∆+=
( ) 





≠
=
=
10
00
x
x
xU

78
where the pair of coordinates, Xi and Yi ,corresponds to the location of the center of the region of
interest in the current frame. On the other hand, the second frame becomes in the current frame,
and the new frame in the sequence of images (third frame) becomes in the next frame. Thus, the
updated location of the region of interest defines the position of the object being tracked.
Finally, the proposed algorithm can be summarized in the following steps:
Determine θ, (Equations 1-6)
Define the value of S
For k=0 to S do
end
Select
Update the location of R1
4. DESIGN OF A SOFTWARE PLATFORM TO CREATE VIDEOS OF
TESTING
To test the performance of the proposed algorithm, it is necessary to create a set of virtual
scenarios where objects with different shapes and sizes wander into scenes through the time.
Therefore, it is designed a software platform that is able to create a wide range of videos where
not only the object to be tracked is present in the scenes but also background objects. These
background objects may be motionless or they may be moved in different directions through the
video. Additionally, the software platform is able to create video sequences under heavily
cluttered conditions, meaning by cluttered conditions the presence of a considerable number of
background objects that interact with the object being tracked. The cluttered conditions constitute
an important challenge for tracking algorithms because the region that surrounds the object of
interest presents a high variability that hinders the modeling of the tracking process in general.
The software platform, which was designed in Matlab 7.12.0, is able to create a wide variety of
scenarios by setting the following parameters:
Number of objects presents in the video. The number of objects which will be located in the
scenes may be defined by the user or may be selected automatically by the software platform.
Shapes and sizes of the objects present in the video. The user may select among a diverse set of
geometric figures the shapes of each object present in the video. Among these geometrical forms,
we have rectangles, triangles, ovals, and other forms composed of two or more of these
geometrical forms.
)sin(*
)cos(*


ky
kx
k
k
=∆
=∆
( )
( )
( )∑=












∆+∆+
−
=∆∆
L
i ii
ii
yyxxR
yxR
UyxM
1
1
,*
,
,
( )
{ } 





∈∀
∆∆∆∆
Sk
yxMMinyx kkkkk
..2,1,0
,,|,

79
( )








≤≤
=
Otherwise
Jx
Jxf
0
0
1
Levels of intensity of each object present in the video. In the same way, the gray-level intensity of
each object located in the video is selected by the user or defined randomly by the software
platform as a value which ranges from 0 to 255.
Location of each object in the video. The coordinates of each background object in the first frame
of the video may be initialized by the user or defined randomly by the testing platform.
Movements adopted by the objects in the video. The background objects may be motionless
through the video or they may be moved around along with the object of interest according to one
out of three strategies, i.e., a random movement, where it is selected randomly two numbers
whose values ranges from 0 to 5 pixels. The two values represent the x-axis and y-axis
displacements of the object in the next frame. The uniform distribution, which is defined by the
Equation (20), is used to generate the displacement values. Once the object has been displaced,
two new random values are obtained to move the object in the next frame. This process is
repeated for the total number of images which constitute the video.
(14)
The parameter, J, in the Equation 14, represents the maximum displacement that the object may
carry out. The second strategy corresponds to a predefined displacement, and it consists in
selecting two random values for the x-and y-axis displacements of the object being tracked in the
same manner as was done for the first strategy of random movement described above. However,
these two values are adopted as fixed displacement of the object of interest between two
consecutive frames, and for all the sequence of images which constitute the video. In other words,
the displacement of the object is always the same between two consecutive images. The last
strategy is a combination of the previous two, where the displacement of the object has a 50
percent chance of being completely random (first strategy) and 50 percent chance of being a
predefined displacement (second strategy).
On the other hand, the frames generated have a size of 600 x 600 pixels, and every video is
constituted by 100 frames. Different geometrical shapes as squares, rectangles, ovals, triangles,
and some combinations of the previous ones were used for the object to be tracked.
5. EXPERIMENTAL RESULTS
The proposed algorithm was implemented in Matlab 7.12.0, and different tests were carried out in
a PC, Dell Inspiron 640 m, with 2 GB of RAM memory. Initially, it is defined a region of interest,
R, in the first frame of the video. This region is placed in such a way that it is completely inside of
the object to be tracked. Once the location of the region of interest has been defined in the first
frame of the video sequences, the proposed algorithm updates automatically the location of this
region of interest for all the remaining video frames. In other words, the location of the object of
interest is automatically identified through the video by the proposed algorithm. Additionally, the
size of the regions of interest used for all the experiments is of 20 x 20 pixels. Finally, the results
obtained by the proposed algorithm with the set of virtual videos are shown form the Fig. 5 to the
Fig. 8.

80
Figure 5: Tracking of an object in the first video sequence
The Fig. 5 shows six particular frames from a video sequence constituted by 150 frames. The
video sequence shown in the Fig. 5 presents a group of object moving in different directions. The
object to be tracked is selected in the first frame by placing a red square-shaped region in the
object. Then, the algorithm automatically updates the position of this square region in such a way
that this region move along with the object through the video sequence allowing the tracking of
the object of interest.
Figure 6: Tracking of an object in the second video sequence
Similarly, the Fig. 6 shows the tracking of an object in a video sequence constitutes by 90 frames. Once
again, the object to be tracked is selected in the first frame, and then the proposed algorithm automatically
locates it in the rest of the video sequence.

81
Figure 7: Tracking of an object in the third video sequence
The Fig. 7 shows another video sequence constituted by a group of objects moving randomly in
time. The object to be tracked is selected and identified by a red square-shaped region that is
placed in the object. The results obtained in this video sequence show that the object of interest is
successfully tracked.
Finally, the Fig. 8 shows the results obtained by applying the proposed algorithm to a video
sequence of 100 images. Each image has approximately 100 background objects that interact with
the object being tracked in different ways. Once again, the object being tracked, which is
identified by a red square-shaped region, is successfully located in the video.

82
The obtained results shows that the proposed algorithm can track several objects in scenarios with
heavily cluttered conditions, where the region which surrounds the object of interest experiences
continuous changes because of the interaction between the object being tracked and the
background objects located in the scenes.
6. CONCLUSIONS
The novel algorithm proposed in this paper to track an object in video sequence was successfully
tested under a wide variety of scenarios where heavily cluttered conditions were present.
Additionally, it was designed a testing platform which allowed to create challenging scenarios
used to test the proposed tracking algorithm. This testing platform was an important tool at the
beginning of the design of the proposed algorithm as well as in the analysis of the tracking
process in general, and it becomes in an important instrument to study in future works other
phenomena presented in video object tracking such as occlusion, scale changes, illumination
changes, etc. On the other hand, the proposed algorithm based on region of interest stayed
immune to heavily cluttered conditions because this algorithm ignores much of the variability in
the environment which surrounds the object being tracked.
REFERENCES
[1] G. L. Foresti, Object Recognition And Tracking For Remote Video Surveillance, IEEE Trans.
Circuits Syst. Video Technol., 9(7):1045-1062, October 1999.
[2] A. J. Lipton, H. Fujiyoshi, R. S. Patil, Moving Target Classification And Tracking From Real-time
Video, Applications of Computer Vision,1998. WACV ’98. Proceedings., Fourth IEEE Workshop on,
pp. 8-14, 1998.
[3] Y. Li, A. Goshtasby, and O. Garcia, Detecting and tracking human faces in videos, Proc. ICPR’00
vol. 1, pg. 807-810 (2000).
[4] Gabriel, P.; Hayet, J.-B.; Piater, J.; Verly, J., “Object Tracking using Color Interest Points,” IEEE
Conference on Advanced Video and Signal Based Surveillance, 2005.

83
[5] C.Harris; M.Stephens., “A combined corner and edge detector,” 4th Alvey Conference, pages 147–
151, 1988.
[6] J.J. Koenderink and A.J. Van Doorn,”Representation of local geometry in the visual system,”
Biological Cybernectics,55(6), 1987.
[7] Haritaoglu, I., Harwood, D., and Davis, L., “real-time surveillance of people and their activities,”
IEEE Trans. Patt. Analy. Mach. Intell. 22, 8, 2000.
[8] Sato, K. and Aggarwal, J., “Temporal spatio-velocity transform and its application to tracking and
interaction,” Comput. Vision Image Understand. 96, 2, 100–128,2004.
[9] Sebastian, P.;Yap Vool Voon,“Tracking using normalized cross correlation and color
space,”International Conference on Intelligence and Advanced system, ICIAS, 2007.
[10] Ashwani Aggarwal, Susmit Biswas, Sandeep Singh, Shamik Sural, and A.K. Majumdar, “ Object
Tracking. Using Background Subtraction and Motion Estimation in MPEG Videos”, ACCV. 2006,
LNCS, vol. 3852, pp. 121-130, Springer, Heidelberg (2006).
[11] D. Comaniciu, V. Ramesh, and P. Meer, “Real-time tracking of non-rigid objects using mean shift,”
IEEE Proc. on Computer Vision and Pattern Recognition, pages 673–678, 2000.
[12] David G. Lowe, ”Distinctive Image Features from Scale-Invariant Keypoints”, Int’l Journal of
Computer Vision, vol. 60, pp. 91-110, 2004.
[13] Khan, Z.H., Gu, I.Y.-H., TieSheng Wang, Backhouse, A., “Joint anisotropic mean shift and consensus
point feature correspondences for object tracking in video,”IEEE International Conference on
Multimedia and Expo, 2009.
[14] Chakraborty, D.; Patra, D.,”Real time object tracking based on segmentation and Kernel based
method,“ IEEE International Conference on Industrial and Information systems, 2010.
[15] R. V. Babu, P. Perez, and P. Bouthemy, “Robust tracking with motion estimation and local kernel-
based color modeling,” Image and Vision Computing, vol. 25, issue 8, pp. 1205-1216, 2007.
[16] Chun-Te Chu; Jenq-Neng Hwang; Hung-I Pai; Kung-Ming Lan, “Robust video object tracking based
on multiple kernels with projected gradients,” International Coference on Acoustics, Speech and
Signal Processing (ICASSP), 2011.
[17] Chun-Te Chu, Jenq-Neng Hwang, Shen-Zheng Wang, Yi-Yuan Chen ,“Human tracking by adaptive
Kalman filtering and multiple kernels tracking with projected gradients,” IEEE International
Conference on Distributed smart Cameras, 2011.
[18] Zhi Liu; Liquan Shen; Zhongmin Han; Zhaoyang Zhang, “A Novel Video Object Tracking Approach
Based on Kernel Density Estimation and Markov Random Field ,“ IEEE International Conference on
image processing, 2007.
[19] Khatoonabadi, S.H.; Bajic, I.V., “Video Object Tracking in the Compressed Domain Using Spatio-
Temporal Markov Random Fields,” IEEE Transaction on Image Processing, 2013.
[20] Amer, A., “Voting-based simultaneous tracking of multiple video objects,” IEEE Transaction on
Circuits and Systems for Video Technology, 2005.
AUTHORS
Andres Alarcon-Ramirez received a PhD degree in Electrical and Computer
Engineering from Howard University. He received his M.S. in Computer Engineering
from University of Puerto Rico (2009), where he was a research assistant for The
Bernard M. Gordon Center for Subsurface Sensing and Imaging Systems. He also
received a M.S. in Electrical Engineering (2006), and his BS in Electrical Engineering
(2003) from Universidad del Valle (Cali-Colombia). Currently, he is working as
research assistant at the Electrical Department at Howard University.
Mohamed F. Chouikha (M '88) received a Ph.D. degree in Electrical Engineering from
the University of Colorado in Boulder in 1988. Since 1988, he has been with Department
of Electrical Engineering at Howard University. In July 2000, he became the Chair of the
EE Department and had since held the position. Dr. Chouikha’s research interests include
multimedia signal processing and communications, and wireless Communications.

A New Algorithm for Tracking Objects in Videos of Cluttered Scenes

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to A New Algorithm for Tracking Objects in Videos of Cluttered Scenes (20)

Recently uploaded (20)

A New Algorithm for Tracking Objects in Videos of Cluttered Scenes