ponent, 
day. In 
ibsolute 
tem.. 
  
14 
acy 
research 
of the 
recision 
metrics, 
tial GPS 
nd CCD 
SPRS94, 
bile 3-D 
Mobile 
995, pp. 
Schwarz, 
ontrolled 
anada. 
‘hapman, 
y Survey 
nference, 
Gong, P. 
to the 
. 59, No. 
STEREO-IMAGE REGISTRATION BASED ON UNIFORM PATCHES 
M. Abbasi-Dezfouli 
Computer Science and Statistics Dept 
College of Science, Shahid Chamran University 
Ahwan, Iran 
T.G. Freeman 
School of Computer Science 
Australian Defence Force Academy 
Canberra, ACT 2600, Australia 
Commission III, Working Group H3002 
KEY WORDS: Registration, stereo-images, image matching, uniform patches, feature-based matching, automatic 
registration, SPOT. 
ABSTRACT: 
The automatic processing of stereo-images to determine terrain height has proved to be a far from trivial problem. 
Methods have been developed by a number of people that rely on operator guidance in establishing ground control 
points or seed points to start the matching process. 
In an attempt to remove the operator interaction altogether, we have developed an unsupervised feature-based 
matching method based on identifying patches of uniform colour. By comparing the shapes and positions of such 
patches, we can establish how coarse features of the two images match, giving us a method for initiating the 
matching process. 
Details of the shape comparison method are presented, along with results when applied to SPOT stereo-pairs and 
close range images. 
1. INTRODUCTION 
The use of stereo-images for estimating distance to 
parts of a scene has been long established in 
photogrammetry and in robotics. The task of 
automating the matching process has been an active 
research area for a decade, with many published 
methods and some commercial systems now available 
(Kauffman and Wood, 1987; Otto and Chau, 1989; Day 
and Muller, 1989; Li, 1991; Schenk et al, 1991; Trinder 
et al, 1990, 1993). Almost all of these methods rely on 
an operator providing seed points to start the matching 
process. 
In an effort to minimize the extent of operator 
involvement in this process, we have looked at the 
challenge of performing the matching entirely 
automatically. If we can succeed at this, it will find 
immediate application in robotics and in close-range 
work with fixed camera positions. Its application to 
photogrammetry will be of less advantage, because 
ground control points are needed to overcome 
uncertainty in the position of the viewpoints and 
viewing direction. Nevertheless, success could reduce 
the dependence in photogrammetry on so many ground 
control points, with fewer requiring to be located in the 
image, and the remainder being supplied as X,Y,Z 
coordinates to be reconciled with the matching process. 
This would alleviate the problem of precisely locating 
all the points in the images, an unfortunate source of 
error in current methods. 
To automate the matching process, we have considered 
the human visual system and the way we would 
perform the task if denied the option of using our eyes 
concurrently, one for each view. Of course, human 
physiology is extremely efficient in performing this 
matching process in everyday life, where both eyes 
concurrently provide the differing views, and the brain 
is able to roughly estimate distance over a whole scene 
within a fraction of a second. The challenge of 
performing similar matching with computer hardware 
remains a very distant goal. 
When humans are denied concurrent access to the two 
views, our visual system is still able to match them, but 
is much less efficient, by several orders of magnitude. 
When confronted with two views, the eyes rapidly 
search for features that can be located in the other 
image, and from coarse details such as these, finer 
comparisons are made, picking up more subtle 
differences and similarities in the images. 
Different types of feature have been investigated by 
various workers seeking to mimic this behaviour within 
a computer. Edge features, or lines where there is a 
major change in light intensity across them, have 
received much attention (Marr and Hildreth, 1980; 
Harallick, 1984; Greenfeld, 1987; Otto and Chau, 1989; 
Schenk et al, 1991). They are not always easy to work 
with, because of difficulties in defining which direction 
the changes in intensity needs to be measured along, 
and in setting a threshold which will give continuity of 
edges while not swamping the matching process with 
too much unwanted detail. Point features, such as 
corners of objects, are particularly attractive because 
they are easy to compare (Moravec, 1977; Forstner, 
1986, 1987; Trinder et al., 1990), and extension to the 
whole area from the found features is quite natural. 
The approach we have adopted is to look for features 
which are areas of uniform colour (Abbasi and 
Freeman, 1994a). This is motivated by looking at 
images of natural terrain, where water bodies and roofs 
of buildings are often the first features identified. Of 
course, these are not the only features that can be 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996