Artif Life Robotics (2003) 7:6-11 9 ISAROB 2003
DOI 10.1007/s10015-003-0230-3
Cheol-taek Kim 9 Ju-Jang Lee
An active contour model for object tracking using the previous contour
Received and accepted: October 16, 2002
Abstract
An active contour model is proposed for object
tracking using prior information. Conventional algorithms
have many problems when applied in object tracking. The
proposed active contour algorithm, a model using an edge
of an adapted color feature, not only modifies the internal
energy function of the conventional algorithm to extend the
search range and reduce the computational burden, but also
modifies the external energy function to reduce the edge
candidates of the object. The algorithm searches normally
and uses dynamic programming to solve the energy minimi-
zation problem. The main drawbacks of a conventional
snake algorithm, i.e., shrinking, a limited search range, sen-
sitivity to outliers, are improved with the proposed algo-
rithm. We illustrate the effectiveness of our scheme using
some tracking examples.
Key words Active contour model - Snake 9 Object track-
ing 9 Dynamic programming
1 Introduction
Since the beginning of computer vision, the detection and
computation of the motion of an object, i.e., object tracking,
in image sequences has been of enormous interest. Because
of increasing computer power, tracking moving objects in
real time has become more important in recent years. There
is a wide range of applications, e.g., video surveillance, au-
tonomous mobile systems, service and cleaning robots, or
systems to assist car drivers.
C. Kim ([]) 9 J.-J. Lee
Department of Electrical Engineering, KAIST, 373-1 Kusong-Dong
Yusong-Gu, Taejon 305-701, Korea
Tel. +82-42-869-5432; Fax +82-42-869-3410
e-mail: [email protected]
This work was presented, in part, at the Seventh International Sympo-
sium on Artificial Life and Robotics, Oita, Japan, January 16-18, 2002
Many algorithms for tracking moving objects have been
discussed in the literature. In particular, because of increas-
ing interest in building autonomous mobile systems, the
detection and tracking of moving objects in a natural scene
are very important and challenging tasks. In the context
of mobile systems, such algorithms cannot be limited to
static cameras. In addition, real-time constraints must be
satisfied. 1~ The task must be solved in a closed loop between
image acquisition and reaction.
In the literature, algorithms for tracking moving objects
can be divided into three classes: optical flow, correlation,
and feature-based approaches. Among the feature-based
approaches, both model-based and data-driven algorithm's
can be found. A widely used data-driven approach is based
on the so-called active contour model. 1
4,11
The active contour model, which was developed as a
useful tool for the segmentation and tracking of rigid or
nonrigid (i.e., deformable) objects, was introduced by Kass
et al. 1 in 1987. They defined snake energies such as internal
energy, image energy, or external energy. The process of
energy minimization could result in successful segmentation
and tracking. They tried to solve the problem of energy
minimization by the variational approach. Amini et all re-
ported that the method can be numerically unstable, and
there is a tendency for points to bunch up at strong positions
on an edge contour. To remedy this, they proposed a dy-
namic programming formulation of order
O(nm3),
which
unfortunately was fairly slow. Williams and Shah 6 intro-
duced a greedy algorithm which was much faster, of order
O(nm),
and achieved comparable performance in both con-
tour-finding and object-tracking. Chun and Shiu 7 intro-
duced an unbiased algorithm, which was a modified version
of the greedy algorithm, for object-tracking. This removed
the bias characteristic of internal energy. Kim s introduced
an image=flow energy to make snaxels roll along the direc-
tion of motion by using the information of image flow.
The main disadvantages of the existing parametric active
contour model are generally owing to their short search
range and heavy computational burden. This article sug-
gests a novel internal energy to extend the search range and
reduce the computational burden. In addition, in cases of
