2004 
jonal 
(2) 
ed to 
ms. 
and 
)blem 
cause 
Due 
n, the 
solve 
adient 
ed to 
nich is 
| from 
bject’s 
ndary. 
enous 
t only 
0 can 
ossible 
image, 
1art of 
erate a 
mages. 
ly will 
various 
out the 
sequent 
1 along 
| object 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
between adjacent images. 
  
First Image 
    
  
  
   
Color features of object 
mm 
     
  
  
  
Color features of background 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
AE 
Removing the segmentation errors and labeling the object by image processing technique 
  
  
  
  
uam aan sume s 
Initial snaxels of ACM 
  
  
Active Contour IModel and Gradient Vector Flow can be used to extract the shape feature 
  
  
Fig.1 Flow chart of object segmentation. 
4.1 Shape-based Matching 
SMA is a kind of invariant shape description, which defines the 
gravity venter of object as the origin and the longest axis of 
object as main axis. The shape of object on the image can be 
reconstructed aas a shape matrix from SMA (Fig.2, Fig.3). By 
comparing the shape matrix B°" with B°, the degree of 
similarity between two objects OT and OS can be calculated. 
Because the dimensions of the matrixes have to be equal, hence 
the dimension of the matrix n and the similarity p (po poss are 
defined by following formulas: 
Lm * | 
n=K max[max d(A, 707), max d(A,Tos)] 4) 
BB) a A SVE ir £C pe (5) 
nxn‘" 
  
zl 
Where d(A,T) = the distance from A to T. 
  
K = the similarity accuracy factor. 
Pl = 
an 
E es RIT 
7 / 
y. / 
/ qMi + / 
Me G 9 
x y 
N. 74 
  
  
  
Fig.2 Object G in the image domain. 
  
M 
  
  
pas 
  
K*d(M, T.) —— — — —— 
  
  
Fig.3 Shape of object G can be reconstructed from SMA. 
The factor K influences the accuracy of similarity; the higher K 
can describe the shape more exhaustively. Considering saving 
the compute time, we set K=3 in this study. By using SMA, 
the similarity between two objects can be calculated, accordingly. 
the object region on the next image can be found too. SMA not 
only finds the most similar object on the next image, but also 
> 
figures out the difference of scaling and rotation between the two 
shapes. 
4.2 Feature Point Tracking 
The shape region can be extracted automatically by preceding 
steps, consequently, the next step is to track the feature point of 
the object. The area-based matching algorithm can match point 
feature precisely, nevertheless, the success of area-based 
matching algorithms highly depends on the invariance of the 
target window and search window. By means of the 
coefficients of SMA, it is posssible to eliminate the shape 
change of the object. Eventually, the center position of the best 
matching block which with highest probability is the tracking 
result. 
The size of target block is 7 by 7 in this research, and the target 
block is given by the result of TDGO (Lue, 1988). TDGO is an 
interest operator, which can find-the most obvious feature point 
in the image. The boundary of search window is decided by 
the object's boundary rectangle on the next image. This study 
selects the Mean Square Error (MSE) to be the objective 
function of area-based matching. MSE measures the 
magnitude of error as a result of two blocks’ comparison, lower