ads. Cross- 
le are then 
sists in the 
t letters or 
ow the eli- 
ildings. Fi- 
. black fea- 
zuish 4 dif- 
: large buil- 
chools, city 
ng in light 
nce letters 
nap image, 
ces just by 
notice that 
vhite) were 
: dark grey 
pixels, and 
and white 
done by a 
olding ope- 
ind divides 
is (opening 
els of these 
surface fea- 
images co- 
have been 
s 1m /pixel. 
is superimposed on the aerial left image. 
First the images are corrected into an epipolar geo- 
metry (see figure 2). Then geometrical relationship 
between both aerial images and the scanned map is 
determined: a polynomial transform of degree 2 is cal- 
culated using manually selected control points. These 
transforms are only used to transfer the vectorial re- 
presentation of the road network extracted from the 
map into both aerial images as presented on figure 3, 
but no registration has been performed directly on the 
images. 
A disparity map is calculated for the complete scene 
with a classical cross-correlation algorithm using a 
square window of 13x13 pixels. The result, as it can 
be seen on figure 4, is very noisy because of the pre- 
sence of numerous hidden parts and because of the 
very large disparity range required for the complete 
scene (80 pixels). — 
The generation of the DTM is then split up into 3 
steps. First disparity is calculated for each crossroad 
of the network. Then disparity is calculated along the 
road sections joining the crossroads in order to vali- 
date the disparity at the crossroads. Finally a dense 
map of disparity is calculated using the validated cross- 
roads. 
3.1 Disparity at the crossroads 
In order to calculate the disparity at a crossroad 
we consider a large window (30x30 pixels) centered on 
each crossroad. A large window is required because 
of the lack of precision on the crossroad localization 
that can be caused by several factors: precision of the 
map itself, displacement during the road extraction 
operation and accuracy of the polynomial transform 
between the map and the aerial images (see figure 4). 
The disparity histogram is calculated on this large 
window and its most significant peaks are considered 
(see figure 5). The peak given by the lowest disparity 
is selected as the ground point representative. Other 
peaks are usually provided by points at the top of 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
g. 4 - Road and crossroads eziracted from the map 
are superimposed on the disparity image. 
Fi 
- 
  
17600 
  
7000 
  
Of pixels 
  
number 
500 
   
  
  
  
e | 
o 26 50 76 700 
disparity 
Fig. 5 - Disparity histogram at a crossroad. 
buildings. The disparity of the selected peak is then 
recorded as the disparity of the crossroad. 
3.2 Disparity along the road sections 
Since crossroad disparity was calculated ‘using only 
local information, we would like to validate this re- 
sult with more global information, in order to remove 
wrong crossroads and to correct wrong values of dis- 
parity. The graph of roads provides the consistency 
required for this validation step.