ORIENTING DIGITAL STEREOPAIRS BY MATCHING FOURIER DESCRIPTORS 
Yi-Hsing Tseng 
Department. of Surveying Engineering 
National Cheng Kung University, Taiwan, R.O.C. 
Commission III 
KEY WORDS: Automation, Orientation, Feature, Matching, Neural, Network, System, Design 
ABSTRACT: 
This paper presents a fully automatic method to reconstruct the relative orientation of a pair of stereo images. The 
operation of this method can be divided into two stages. First, by using the conception of feature-based matching, 
the Fourier descriptors of conjugate boundaries of homogeneous regions between the images are determined through 
a neural network system. This provides a reliable solution of approximate orientation. Second, the more accurate 
conjugate points are matched by using, template based on the initial orientation. The relative orientation then is 
determined by using the photo coordinates of the matched conjugate points. 
1. INTRODUCTION 
The idea of fully automatic relative orientation has 
been proposed by Schenk, Li and Toth [1991]. They 
suggested to solve the initial orientation by matching, 
edges in y-s domain. Then accurate orientation can be 
determined by using template matching, such as cross- 
correlation or least-squares matching. They also 
indicated that matching entire edges as opposed to 
more traditional point matching methods substantially 
increases the robustness of the solution. However, they 
also commented that it is more difficult to implement 
and there is still room for improvement. 
Based alone on the factor of shape similarity in 
matching features, it tends to obtain wrong, matches, if 
some features in an image are similar in shape. This 
problem can only be solved by taking the account of 
orientation consistency of matched features. Based on 
the idea of feature-based matching, and considering the 
criterion of orientation consistency, a neural network 
system is implemented to obtain the optimal matching 
of the conjugate features of the stereo images. 
The automatic process of relative orientation can be 
divided into two tasks (Fig. 1). The first task performs 
feature-based matching to solve the orientation 
approximately. The second task performs point 
matching to determine conjugate points accurately. 
The relative orientation then can be computed by using 
the photo coordinates of the matched conjugate points. 
The proposed system for feature-based matching 
conceptually mimics the human recognition process. 
Conjugate features are determined not only by the 
condition of shape similarity but also by the fitness of 
orientation consistency. First, homogeneous regions 
are extracted from images by using region-growing 
method. Then, their boundaries are described by using 
Fourier descriptors. By applying the least-squares 
approach to matching Fourier descriptors [Tseng and 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
Schenk, 1992], one can compute the shape similarity 
and relative orientation of matching features. All 
matching conditions will be combined into a cost 
function which can be applied by the Hopfield-Tank 
neural networks in determining the optimal matching, 
No initial values of orientation between images are 
required and it is expected to be adaptive to the 
disturbances of image distortion and noises as well as 
the differences of image orientation and scale. 
de AT pu ELI 
/ Left image of the / Right image of the 
/ coarsest level / / coarsest level / 
6i E vara. JJ EA 
EEE ee E see ies 
f | { 4 
Segmentation of | | Segmentation of : 
| homogenous area features | homogenous area features | 
M un ub 8 MC ALI. I. 
| Computing Fourier | | Computing Fourier | 
descriptors of the feature | | descriptors of the feature | 
boundaries | | boundaries | 
PL rai C ; 
] Task: 
Matching Fourier descriptors| Determining 
approximation 
A neural network approach to 
finding conjugate features 
J 
  
Y 
—— rr 
| Solving approximate | 
orientation | 
THery = Nn 
A 
TTT x | 1. 7^ — | 
Finding interest points i ty Finding conjugate points | 
| left image ) in right image ! 
S fs i = EN AS 
ouo Gui | 
Fu ED Task M: 
| Coarse-to-fine matching | Point 
| process | matching 
an 22 
| 
| Bundle adjustment | i 
| Nr e iE 
Figure 1: The work flow of the system 
For the task of point matching, interest points are 
selected on the left image. Then, the initial conjugate 
      
   
   
   
   
    
   
     
    
   
   
   
     
   
   
   
  
    
     
   
   
    
   
   
    
    
   
    
   
      
   
  
IT 
n 
(D + “0° OTE e) €. Fr 
Un (T E EM 
as