International Journal of Engineering and Advanced Technology (IJEAT)

ISSN: 2249 8958, Volume-1, Issue-3, February 2012

Tracking the human motion in real time using

Star Skeleton Model
B. MUTHUKUMAR, S.RAVI
Abstract: Human motion analysis is receiving increasing

In general, they work by detecting features (such

attention from researchers. This interest is motivated by wide

as hands, feet and head), tracking them, and fitting them to

spectrum of applications. In this paper, a process is described

some a priori human model.

for detecting moving targets and extracting boundaries. From

There are two main drawbacks of these systems:

these, star skeleton is produced. The star skeletonization is

they are completely human specific, and they require a

suitable for detecting and analyzing human motion in real time.

great deal of image-based information in order to work

Also the method does not require great deal of image-based

information to work efficiently. Extremal points are extracted

effectively. For general video applications, it may be

in star skeleton like head, hands and legs, their tracking

necessary to derive motion analysis tools which are not

described based on an n*n block of DCTs coefficient. Then we

constrained to human models, but are applicable to other

correct the false tracked extremal points such as occluded

types of targets, or even to classifying targets into different

extremal points.

types. In some real video applications, such as outdoor

Index Terms: Human Detection, Image Processing, Occlusion

surveillance, it is unlikely that there will be enough pixels

Removal

on target to adequately apply these methods. What is

I. INTRODUCTION

required is a fast, robust system which can make broad

One of the more active areas is activity

assumptions about target motion from small amounts of

understanding from image sequence. Understanding

image data.

activities involves being able to detect and classify targets

This paper describes the star skeletonization

of interest and analyze what they are doing. Human motion

with correction method for analyzing the motion of human

analysis is one such research area. There have been several

targets. The notion is that a simple form of skeletonization,

good human detection schemes, which use static image

which only extracts the broad internal motion features of a

sequences. But detecting and analyzing human motion in

target, can be employed to analyze its motion. We use

real time from video image sequence has only recently

Discrete Cosine Transform (DCT) for tracking and

become viable with algorithms like Pfinder [3] .These

matching algorithms which are based on Reference Matrix

algorithms represent a good first step to the problem of

Descriptors (RMDs) [2]. The 2D tracking process is

recognizing and analyzing humans, but they still have

performed by DCT block it corrects the false tracked

some drawbacks.

points. In section II we review the human Star

____________________________________________

Skeletonization for reconstruction. In section III we give

B.MUTHUKUMAR, Associate Professor, Department of Information

an overview of algorithm for tracking and matching using

Technology, Kathir College of Engineering, Coimbatore -62. Tamilnadu,

India

(9566907501)

E-mail:

DCT matrixes and method to detect and correct occluded

[email protected]

joints. Finally section IV is conclusion.

Dr. S.Ravi, Professor, Department of ECE, Dr. MGR University,

Chennai Tamilnadu, India (9840148981) E-mail: [email protected]

87

Tracking the human motion in real time using Star Skeleton Model
II. HUMAN DETECTION
A. Background Subtraction
For real-time target extraction we use the
approach background subtraction. Background subtraction
provides the most complete feature data, but is extremely
sensitive to dynamic scene changes due to lighting and
extraneous events. Background image is not fix and it must
adapt to Illumination changes, motion changes like tree
branch move and changes in background because of
objects entering in scene and stays for longer period
without motion.
The basic method is modeling background as

Fig 1: Target pre-processing. A moving target region is morphologically

color image. Background is estimated as the average or the

dilated then eroded. Then its border is extracted.

median of the previous n frames. If the distance between

current frame and background is greater than some

C. Star skeletonization

threshold for some pixel then that is considered as

An important cue in determining the internal

foreground.

motion of a moving target is the change in its boundary

B. Target pre-processing

shape over time and a good way to quantify this is to use

For Star skeletonization we need object contour.

skeletonization.

No motion detection algorithm is perfect. There will be

spurious pixels detected, holes in moving features.
Foreground regions are initially filtered for size to remove
spurious features, and then the remaining targets are preprocessed before motion analysis is performed.
Foreground

region extraction

may produce

images that contain holes and gabs in the object detected,

and may cause objects to be split into more than one
connected region, which would make the object to be
detected as multiple objects. So we need to apply sequence
of dilations and one erosion [6].
Fig 2: Star Skeletonization

The method proposed here provides a simple,

real-time, robust way of detecting extremal points on the
boundary of the target to produce a star skeleton [1]. The
star skeleton consists of only the gross extremities of the
target joined to its centroid in a star fashion

88

International Journal of Engineering and Advanced Technology (IJEAT)

ISSN: 2249 8958, Volume-1, Issue-3, February 2012
III. TRACKING WITH ERROR CORRECTION

We use this three frequency region for tracking and

We locate head, hand and leg like extremal points

matching algorithm. Matrix distance: in special frequency

position using uncalibrated monocular video and use them

region is defined according to the following equation:

to estimate star skeletal configuration. In the proposed

algorithm, head, hand and leg tracking is performed based
on an n*n block of DCTs coefficient (descriptor matrix).
Then we correct the false tracked extremal points such as
2) Reference Descriptor Matrix (RDM)

occluded extremal.

These matrices save tracked extremal points

Algorithm starts with background subtraction and

information of previous frames and database information

then star skeletonization to specify head, hand and leg like

and used to find Corresponding extremal points in tracking

extremal points positions in initial frame. Then for each

process. We generate a reference descriptor matrix for

extremal point, descriptor matrix is computed and saved as

each extremal point (RDM1 ...RDM5).

Reference Descriptor Matrix" for the same point. In the

Reference descriptor matrix for extremal point j

next stage, all extremal points are tracked with its own

(RDMj) is loaded from descriptor matrix for extremal

RDM and are updated based on frequency regions.

points j after initialization and updated after finding new

After finding extremal points positions in next

tracked extremal point j and correction in next frame.

frame, the correction algorithm will correct the primal

Updating routine is different in each frequency region

extremal point position.

3) Tracking

A. Joint Tracking Process

The tracking process is based on frequency

Tracking and database matching is discussed

domain matching techniques. Tracking process aims to

briefly in this section.

find extremal point in successive frames search window

1) Descriptor Matrix

(SWDMs). Finding best match is performed by selecting

We use DCT block for tracking and matching

minimum matrix distance between low and middle

purposes. "Descriptor Matrix"[4] for the point Pi is n*n

frequency of RDMj and search window descriptor matrices

DCT coefficients matrix. By putting the image window of

(SWDMs). We consider best matching point in the search

fixed size (n*n) centered on point Pi into matrix A, a

window as tracked extremal points at frame t+1 and save

descriptor matrix (DM) for Pi is then compute by

its matrix distance named "Mdis" as a parameter to

determining the best tracked extremal points. These
extremal points are initial tracked joints at frame t+1 and
will be used as "start points.
4) Database matching process
Where px and py are pi coordinates. Also, if x=0;

The database consists of required information for

C(x) =1/n otherwise C(x) = 2/n There are n2 coefficients

different poses of video sequences of number of subjects.

Measuring similarity between human pose in current frame

in each DM matrix divided into three frequency regions.

89

Tracking the human motion in real time using Star Skeleton Model

Fig.3 Occlusion Correction Algorithm

(Pf) and human pose in database (requires two kinds of

minimum energy. Energy function is constructed to correct

matrix: DDMs and FDMs, which will be defined later. If

tracked joint location according to global pose of body.

pose distance is smaller than , corresponding occurs;

IV. CONCLUSIONS
In this paper, a method for human detection via

therefore points and middle frequency of RDM must be

star skeleton model is described. This system uses an

corrected. Pose distance is defined by:

adaptive background model robust against long-term

changes, illumination changes and repetitive clutter motion
in the scene. A method for solving temporal occlusion
problem using energy functions is described. In this
method, extremal point tracking is performed based on an
Database descriptor matrix (DDM) is generated

n*n block of DCTs coefficient (descriptor matrix). Then

using low frequency of RDM (for extremal points

we correct the false tracked joint such as occluded joint

similarity) and middle frequency of database (for edge

using an energy function which can estimate false joint

similarity)

position.

REFERENCES
[1] Fujiyoshi, A. J.Lipton and T. Kanade "Real-time Human Motion
Analysis by Image Skeletonization" IEICE TRANS, 2004.
[2] N. Roudsarabi. and A. R. Behrad, "3D Human Motion Reconstruction
Using DCT Matrix Descriptor", ICISP 2008, Vol. LNCS 5099, pp: 386-

B. Correction
After tracking process, we have primal location of

395, 2008.
[3] Wren, A. Azarbayejani, T. Darrell, and A. Paul Pentland "Pfinder:

all extremal points. The correcting process is based on the

Real- Time Tracking of the Human Body" IEEE Transactions on Pattern

minimization of energy function. Its basic idea is to correct

Analysis and Machine Intelligence, 1997.

extremal points position in each frame. Energy function is

[4] Nadiya Roudsarabi, Ali Reza Behrad, Solving Occlusion Problem in

3D Human Motion Reconstruction 2008 International Symposium on

computed for every pixel in the search window (SW).

Telecommunication.

Finding best match is performed by selecting point with

90

International Journal of Engineering and Advanced Technology (IJEAT)

ISSN: 2249 8958, Volume-1, Issue-3, February 2012
[5] Weiming Hu, Tieniu Tan, A Survey on Visual Surveillance of Object
Motion and Behaviors IEEE transactions on systems, man, and
cyberneticspart c: applications and reviews, vol. 34, no. 3, august 2004
[6] B. Jahne "Digital Image Processing, Concepts, Algorithms and
Scientific Applications" 4th edition, 1997.

AUTHORS BIOGRAPHY
B.MUTHUKUMAR received B.E (CSE) degree from
Anna University, Chennai in the year 2005.M.Tech
(CSE) with Gold Medal from Dr. MGR. University in
the year 2007, and Distinction with Honors. Now
Pursuing Ph.D in St. Peters University, Chennai from
2008. Having 6 Years of Teaching Experience. His area of interest in
Image Processing and Networks. He published 07 International Journal,
15 International Conference and 15 National Conference and 25 National
Workshops/FDP/Seminars etc., He is a member of ISTE, CSI, IEEE,
Member of IACSIT and Member of IAENG.

Dr.S.RAVI received DEEE, A.M.I.E (ECE), and M.E

(CS) degree in the year 1989, 1991 and 1994
respectively and the Doctorate degree from Anna
University, Chennai in the year 2004. He is having 16
years of teaching experience in engineering college. His areas of interests
are VLSI, Embedded Systems, Image Processing and Simulation and
Modeling. He has published more than 40 papers in International /
National Conferences and 30 in Journals. He is a member of IEEE and
ISTE and Life member of IETE and IETUK.

91