n of the 
juired. 
y, either 
or by 
ages. 
CH 
lation is 
] which 
d. The 
As the 
d by the 
5 source 
nd, it is 
appears 
scted or 
e height 
es, thus 
ature on 
the left 
spacing 
height 
use in a 
window 
user. 
splayed 
v of the 
which 
purpose 
s of the 
ung the 
ors, the 
Control points 
existing DEM 
   
Moravec operator 
   
Least square 
adjustment 
linear features 
matching 
   
  
  
gridded DEM 
Delaunay 
riangulatio 
orthoimages 
Figure 2. 
position of the icons and the windows on the screen are 
gathered in one configuration file whose syntax is 
understandable by the operator : keyword = value. 
The GCPs are measured manually on a screen, based on 
enlargements of small windows of the two images and the 
orientation subsequently derived, according to the formulae 
described in Section 1.1. The GCPs are then triangulated, 
and the user can display this triangulation as an overlay. 
Numerous possibilities are then available, e.g.: zoom in on 
any area at any scale; or display any area in stereo in two 
separate subwindows at true resolution and at five times the 
true resolution to precisely identify features (road crossings, 
river forks, buildings etc.). The user can thus select an area 
where network densification is required and start the 
automatic matching process. 
There are several procedures to allow the operator to check 
and control the generation of the DEM. Firstly, the user can 
display some areas in stereo and check if the points added are 
truly matching. He/she can, as well, display a perspective 
view to test the shape of the model. Left and right ortho- 
images can also be generated so that the user can check 
visually if they correspond, and if not, identify the areas 
which have been incorrectly matched. It is then possible to 
alter some of the points of the triangulation or even to 
suppress them. 
As all these operations are highly interactive it is vital to have 
a system which responds quickly to any user request. The 
program must be carefully designed to avoid serious 
operational problems for the user. 
3. IMAGE MATCHING 
The image matching is based on a sequential process of 
feature matching followed by area based matching. Each of 
the steps will be described below. 
3.1 SELECTION AND MATCHING OF 
FEATURES 
The feature matching of SPOT images is based on point 
features or 'interest points' after Moravec (1977). The 
971 
   
  
perspective views 
extraction of the interest points from images, according to 
Moravec method is undertaken in two steps: 
. For every pixel in both the left and right image a 
value of a so-called interest operator is calculated. 
. Thresholds of the values of the interest operator for 
both images are chosen, and those pixels above 
threshold in both images are qualified as interest 
points. 
In the case of the Moravec method, the interest operator is the 
minimum of the slopes in intensity derived in four directions. 
The evaluation of the interest operator commonly uses a 5x5 
window around each pixel which is a very time consuming 
task. In addition, the determination of thresholds for the 
interest operator for whole images can often be difficult, 
especially in the case of highly unhomogeneous images. 
Matching of interest points selected in both images may 
typically be based on criteria such as those described by 
Barnard and Thompson (1981). That is, the matched interest 
points must have similar characteristics of discreteness, 
similarity, and consistency. Discreteness is determined by the 
interest operator. Similarity is determined by the similarity of 
the intensity patterns of the potentially matching interest 
points on both images. Based on this test, the decision is 
made as to which of all possible combinations of interest 
points on the two images may potentially match by virtue of 
the intensity distributions around them. The final test of 
consistency aims to eliminate erroneous matchings by testing 
the consistency of the geometry of the sets of possible 
matching points on the two images. This can be done 
typically by an affine transformation of the positions of 
potential matched points in each image. 
In order to avoid the need to perform the very time consuming 
task of generating the so-called Moravec image, the method is 
modified so that the interest points are selected in a set of 
small sub-areas distributed over an image. A sub-area is 
typically 30x30 pixels in size and its centre is positioned in 
the left image by arbitrarily choosing points such as nodes of 
a regular grid or of the triangular network. The corresponding 
centres of the sub-areas in the right image are predicted using 
a priori information about the stereo-model and the 
topography of the terrain itself (roughness of the terrain and