ne level.
1 strategy
Will look
1g to the
1 easy-to-
requires
lity and
eliminary
ach road
utes. To
e tokens
ection in
liability
tions by
on begins
on these
ze is Very
- and the
sach road
le road
le proper
rameters.
ollowing
ng based
the "top-
ble road
itiate the
action of
The top-
order to
tiate the
rs to be
ttributes :
tive class:
oad
f lanes: 2
' separated
1
the road:
type : rural
eadjustment
black).
road seed.
Figure 4 : The map, an image understanding tool
6. MAP AIDED INTERPRETATION
The topographical map which is used as external data
source is a 1:25 000 scale scanned document which
content is quite similar to the final result (BD Topo®)
that is intended to be acquired. The symbolization of the
geographical objects is directed by the legend and by the
map drawing rules, enabling a real representation
homogeneity. The implicit relations existing between the
different objects are expressed by the visual variables :
shape (geometric, figurative or symbolic), size, color,
orientation, The characteristics of these variables
(associativity, differentiation, order and quantity) and
their arrangements define the cartographic language, the
semiology. It provides a potential global comprehension
of the scene (cf Figure 4), emphasizing the networks
organization and the spatial relationships existing
between the different objects layers.
In a first time, the scanned map is digitized : the road
network and the other cartographic objects are identified
and reconstructed to be available for the image
understanding task (Guerin, 95a).
As all information issued from a treatment, the map is
subject to unavoidable distortions with regard to the
ground reality. These alterations have consequences at the
geometrical, topological and semantic levels. The
preliminary study of the map quality (Guerin, 95b) has
outlined the main tendencies of the distortion model,
especially for the road network :
e from a quantitative viewpoint : maximum values of
distortion, general statistics on a test set.
e froma qualitative viewpoint : role and importance of
the context on the distortions, local configuration and
information density influence, topological
inconsistencies detection and study.
The results lead to a prediction model of the road
network aspect.
This study has shown that in these topographical maps
the global topology and geometry of the road network is
correct except near complex cross-roads and dense areas.
The accuracy of the document allows to use a method of
readjustment, implemented as follows : the road network
extracted from the map is registered on the image and
distorted in order to fit the roadways detected in their
neighbourhood. By this way, the image analysis problem
is restricted to a matching process.
Two operators are developped, one specialist for the
detection and the restitution of the intersections, another
for the roadways. Using the map knowledge, the
crossroads defined as reliable by the prediction model are
first positionned. They carry valuable information
describing their local shape, the average radiometry of the
surface, the position and orientation of the incoming
roads. Then the roadways are sought, starting from the
validated cross-roads. By this way, the detections are
propagated all along the graph of the road network.
a estimula fi >
Sans
ll nl
Figure 5 : Focus mechanism, allowed by the external
data knowledge
7 CONCLUSION
The use of external data is a promising solution for image
understanding. But through the two examples given in
this article we can see that the way this external data can
be used is closely linked to the content, the scale and the
accuracy of the data source.
In part 5, the cartographic database is very rich
semantically but the geometry and the topology of the
road network are not reliable, thus the road extraction
method is very local and have to be adjusted for each road
section, according to its reliability and to its
characteristics.
137
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996