left image 
   
right image 
Fig. 2 - Portion of the stereo pair of aerial images. 
2.1 Road network 
On IGN ! maps at the scale of 1:25000 main roads 
are represented in red, while other roads remain in 
white. A simple colorimetric analysis allows the ex- 
traction of the main roads. A morphological opening 
removes contour lines represented with the same color, 
but with thinner lines. 
Other roads are represented with two parallel black 
lines. Their extraction is performed using the road net- 
work connexity. Linear elements with a high probabi- 
lity to be road portions (ie., white pixels bordered 
with black pixels) are detected on the scanned map. 
Links between these elements are created if it is pos- 
sible to find a path from one to an other composed 
of only white pixels (at the crossroad level it is not 
possible to use the border to follow a road). 
Both networks are merged, and only large connex 
components are kept in order to remove elements de- 
tected between black objects representing buildings. 
A linear approximation of the network gives a vecto- 
rial representation. Collinear segments are grouped to 
form roads. 
At that point, an interactive tool allows us to cor- 
rect the road network automatically detected. A small 
part of the errors are some lines detected inside the 
building blocks. Other errors correspond to missing 
road parts due to letters or city limit lanes overlap- 
ping the roads on the paper map. 
Crossroads are detected by grouping the road junc- 
tions (X, T or L shaped junctions): a crossroad could 
be defined as only one junction, or could gather se- 
veral junctions if they are close enough. Each road is 
then decomposed into sections: a section is the portion 
1. IGN: Institut Géographique National, the French na- 
tional agency for cartography. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
of a road between two consecutive crossroads. Cross- 
roads and sections which are not in a cycle are then 
eliminated. 
2.2 Word extraction and urban block 
classification 
The first step for word extraction consists in the 
detection of black features, which represent letters or 
large buildings. Size and shape criteria allow the eli- 
mination of some elements representing buildings. Fi- 
nally the grouping of horizontally aligned black fea- 
tures gives the words present on the map. 
Except roads and letters, one can distinguish 4 dif- 
ferent surface features on the scanned map: large buil- 
dings in black, administrative buildings (schools, city 
halls, ...) in dark grey, apartment building in light 
grey and building-free surfaces in white. Once letters 
and roads have been removed from the map image, 
it is possible to separate the different surfaces just by 
considering the pixel grey levels. One can notice that 
only 3 different tones (black, grey and white) were 
used for the representation of these regions: dark grey 
features are composed of black and grey pixels, and 
light grey features are composed of grey and white 
pixels. The separation of these 3 tones is done by a 
k-means algorithm that performs a thresholding ope- 
ration on the image grey level histogram and divides 
it into 3 classes. Morphological operations (opening 
and closing) allow the grouping of the pixels of these 
3 classes to form the 4 different kinds of surface fea- 
tures. 
3 Digital Terrain Model 
For this application a pair of stereoscopic images co- 
vering the same scene as the scanned map have been 
acquired. The resolution of both images is 1m/pixel. 
698 
Firs 
met 
betv 
dete 
cula 
tran 
pres 
map 
but 
ima, 
wit 
sque 
bes 
senc 
very 
scen 
step 
of t] 
roac 
date 
map 
roac 
3.1 
h 
we c 
each 
of t. 
that 
map 
oper 
betv 
Win: 
(see 
is se 
peal