DEVELOPMENT OF EFFECTIVE PROCEDURES 
FOR AUTOMATIC STEREO MATCHING 
Yu.V. Vizilter*, S. Yu. Zheltov*, 
*State Research Institute of Aviation Systems, Moscow, Russia 
7, Victorenko str., Moscow, Russia, 125319; zhl@gosniias.ru 
KEY WORDS: Photogrammetry, Stereoscopic, Matching, Correlation, Information 
ABSTRACT: 
It is well known that the correlation stereo matching algorithms are the basis of most of digital photogrammetric systems. This paper 
concerns the important problem of increasing of computational speed of automatic stereo matching based on elimination of image 
parts with low information presence and non-traditional implementation of computations using the “sliding window” technique. 
1. INTRODUCTION 
It is well known that the correlation stereo matching algorithms 
are the basis of most of digital photogrammetric systems. This 
paper concerns the important problem of increasing of 
computational speed of automatic stereo matching based on 
image informative characteristics and non-traditional 
implementation of computations in a sliding correlation 
window. 
From our experience of operation with well known Russian TK- 
350 spaceborn images (Sibiryakov, 2000), they usually include 
some low-informative regions. Stereo matching of such regions 
is just a waste of time. Therefore there is a problem of selection 
of sample regions from the viewpoint of their 
selfdescriptiveness under the criterion of accuracy and 
probability of correct stereo matching. At use of correlation 
methods, a correlation coefficient is the best characteristic of a 
signal level inside template. The disadvantage of this 
characteristic that it is calculated during matching, while the 
index of signal level should be calculated before matching, 
indicating on those templates, which will have accurate 
matching. In the developed method an adaptive statistics of 
template brightness is used as an a priori evaluation of 
informative index, which is similar to the correlation 
coefficient. An advantage of this index is that it is calculated 
prior to the beginning of matching. This problem is considered 
in the section 2 of this paper. 
Then the problem of reduction of computational time is 
considered concerning the automatic stereo matching based on 
the normalized correlation. The idea of acceleration of 
calculation of the convolution sums in a sliding window is well 
known (Huang, 1983). The essence of this idea consists of 
storing the ready-made column sums and recursive subtracting 
and adding of the appropriate partial sums corresponding to the 
motion of a sliding window along the image row. It allows 
essentially reducing the amount of calculations in case of usual 
correlation convolutions, but not in case of stereo matching with 
separate convolution fields for all image points. In this work it 
is offered to bypass this problem by means of changing the 
order of calculations. It is proposed to implement the loop by 
disparities as an external loop, and the convolution loop as an 
internal one. In this case all calculations can be implemented in 
a manner of sliding window algorithm, and we obtain the 
required gain in productivity. This problem is considered in the 
section 3 of this paper. 
2. ANALYSIS OF SELFDESCRIPTIVENESS OF 
IMAGE FRAGMENT 
The proposed method for elimination of “empty places” is 
based on analysis of brightness statistical properties of the 
special “wedge” (Figure 1) also captured by this camera. The 
“wedge” here is an image with smoothed intensity changing 
from left to right border. Analysis method contains the 
following. 
The smoothed functions of intensity B(d) and MSE of intensity 
D(d) along the “wedge” are obtained. 
Then the dependence of noise MSE on intensity D(b) is 
estimated. For this purpose one should create the inverse 
function d=B"(b) using the linear interpolation. Then the 
function has a form 
D(b) = D'(B'(b)). (1) 
shown on Figure 2. 
Function O^ — D(b) allows testing the signal presence in the 
image fragment. 
  
  
  
  
Figure 1. The optical “wedge” image. 
Let f; — means the value of signal samples inside the fragment, 
i=1,...,(2N+1)(2N+1). It is required to test the hypothesis Ho 
concerning the data set is homogeneous: 
Ho: f,=u+tn;, n; € N(0,0?(u)), 
where u is the supposed constant brightness value on a 
fragment; n; — noise samples, O(u) -dependence of MSE on 
intensity, derived from the “wedge”. 
   
   
  
  
    
  
   
  
    
   
  
  
  
   
   
   
   
    
    
   
   
    
        
    
     
  
  
  
     
   
  
  
     
    
  
  
  
   
     
    
  
      
Interi 
  
bon D 
For v 
it is 
distri 
to the 
The « 
Then 
prese 
then 
For ] 
form