International Archive 
3. OBJECT SEGMENTATION 
In this section, the detail of object segmentation methods will be 
presented. Dissimilar to the tracking object on the unchanging 
background, it is difficult to use simple differentiating method to 
segment the object on the varied background. Therefore, we 
use two-stage segmentation procedure to track the object on a 
sequence of images. The first stage uses the color features to 
segment possible regions of the target object from the images. 
The second stage employs Active Contour Model to describe the 
contour shape of the target object. 
3.1 The Color Feature Segmentation 
[n the early perception stages of human beings, similar colors 
are always grouped together for further analysis. Based on this 
assumption, desired object can be identified by extracting tne 
characteristic colors as the features. In this study, the desired 
object is manually selected by means of marking a rough 
polygon in the very first image. Then an unsupervised 
classification method is used to extract major color classes of 
the object and background. Accordingly, these color classes 
can be used to segment the object or background on the next 
images. The followings are the brief descriptions of the 
method. 
For a given pixel P is one of pixels in the search window on next 
image. {O,} is the class centers of the object and {B.} is the class 
centers of background. T is the threshold of spectrum distance. 
If (P.Minimum Distance to {B.}<T) 
P belongs to background class. 
If (P.Minimum Distance to {O}<T) 
P belongs to object class. 
Else 
P is a part of background class. 
After the color-based segmentation is finished, it is possible to 
obtain a binary image with object separated from background. 
Normally, the desired object segment should be a solid region; 
therefore, the erosion, dilation and eight-neighbor object 
labeling algorithms are employed to eliminate*the error pixels. 
Moreover, it is necessary to describe the object as a close contour, 
an energy-minimizing edge segmentation algorithm, which is 
applied to represent the shape of the object, will show in next 
sub-section. 
3.2 Shape Feature Extraction 
Active Contour Model (ACM) (Kass, 1987) is one kind of 
parametric curves presentations, which is defined within a curve 
domain. The curves of ACM can move under the influence of 
internal forces cause by the initial curve and external forces 
supplied by image data. ACM is widely used in image 
processing applications, such as edge detection, segmentation 
and particularly to locate object boundaries. ACM transfers the 
boundary detection problem in image domain to 
    
s of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
970 
energy-minimizing problem in curve domain. The traditional 
energy functions in ACM are defined as follows: 
N 
E = Ss: (E. (s;) i EF (s, ) (1) 
i=l 
E.(S) = als, = S + BS... 25; + S d : (2) 
E s (S;) = —[Vf(S, ) (3) 
Where E die = the snake energy of the contour. 
S i = the ith position of contour. 
N = number of snaxel. 
E, 
int the internal energy at snaxel S; . 
E eu = the external energy at snaxel S; . 
a, f! = the weighting functions are defined to 
control the relative importance of the elastic and bending terms. 
Traditional ACM has two problems, initialization and 
convergence in concave regions. The initialization problem 
means that the initial contour has to close to the object, because 
the potential force of traditional ACM is generally small. Due 
to the ACM has no extra pressure force on concave region, the 
contour is often across the boundary concave. In order to solve 
the problems, an extra external force, which is named Gradient 
Vector Flow (GVF) (Xu and Prince, 1998) can be used to 
improve the result of ACM. GVF is a diffusive filed, which is 
computed by gradient vector of a gray-level map derived from 
the image. When the point in the field is near to object's 
boundary, the GVF field will move toward the boundary. 
Moreover, GVF field will change smoothly in the homogenous 
region of the image. Therefore, it indicates that GVF not only 
can provide a larger buffer in initial contour but also can 
converge to the concave region. 
By using GVF Active Contour Model, the shape of possible 
objects, which are segmented by color features on next image, 
can be extracted automatically. Fig.1 shows the flowchart of 
object segmentation. 
4. SHAPE-BASED OBJECT MATCHING AND 
TRACKING 
It must be noted that the color segmentation may generate a 
number of possible objects on the subsequent images. 
Moreover, the movement of the video camera definitely will 
keep changing the shape of the object. Apparently both various 
possibilities and varied shapes of the objects will bring about the 
complexity for tracking the desired object on the subsequent 
images. This study uses the shape derived from ACM along 
with Shape Matrix Algorithm (SMA) to trace the desired object 
  
   
     
  
   
  
  
  
   
  
  
  
  
   
   
   
  
    
  
    
  
  
  
   
   
   
   
  
    
    
  
   
  
  
   
    
  
   
  
    
    
   
    
    
   
  
  
re 
CO 
sil 
Be 
the 
de