This preliminary study is essential and leads to an
adapted analysis method.
Parts 5 and 6 present two examples of external data
contribution to road extraction in aerial images. In the
first case, the external data are provided by a cartographic
database ; in the second one, they are provided by a
scanned topographic map.
5. ROAD EXTRACTION GUIDED BY A
CARTOGRAPHIC DATABASE
In this research, the source of external information is the
IGN Cartographic DataBase (CDB) which contains the
major road network, other networks such as railways and
land cover areas. This database has been acquired by
digitizing 1:50 000 maps and is devoted to the drawing
of 1:100 000 maps. Its geometric accuracy is about 20 m.
The road network, like the other objects of the database,
has semantic attributes which characterize the road aspect
(number of lanes, administrative class...). In this work,
we limit ourselves to the problem of road extraction in
high resolution aerial images.
Given the CDB characteristics and the aerial image
resolution the process which has been chosen is an
algorithm guiding one.
In the preliminary study about the CDB road network
quality (Bordes, 1995), it appears that some objects of
the database are very reliable geometrically, and others are
far less reliable. In a first approximation, this reliability
only depends on the road type and on the road context.
The road network distortions cannot be modelized,
therefore it is impossible to define an overall detection
Figure 1: The CDB road network projected on the
image.
in white : the most reliable road segments
in black : the unreliable road segments
136
method using the CDB everywhere at the same level.
This remark leads us to define an interpretation strategy
so called "easiest at first", that is to say that we will look
in the image for the road sections corresponding to the
database ones beginning by the well-situated and easy-to
see sections. This hierarchic interpretation requires
knowledge about the road sections reliability and
legibility in the image. The results of the preliminary
study allow us to predict the reliability of each road
section location knowing its semantic attributes. To
complete this prediction, we compute image tokens
which confirm or not the location of the road section in
the image. These image tokens and the a priori reliability
of road sections are used to classify the sections by
reliability order (cf Figure E). The road extraction begins
by the most reliable road sections and leans on these
extracted sections to extract less reliable ones.
The second stage in which the external knowledge is very
useful is the choice of the proper road extractor and the
control of its parameters. More precisely, for each road
section, the CDB knowledge about the road
characteristics and context are used to select the proper
road extraction algorithm and to control its parameters.
We use three road extraction methods : a road following
based on homogeneity criterion, a road following based
on profile analysis and a road detection based on the "top-
hat" morphological operator. The most reliable road
sections of the CDB are used as road seeds to initiate the
road followings (cf Figure 3). Then, the extraction of
other roads leans on these reliable road sections. The top-
hat operator is used for unreliable sections in order to
compute a coarse detection which allows to initiate the
road followings. This hierarchic strategy appears to be
efficient for the detection of most evident roads.
| Semantic attributes :
Administrative class:
: Number of lanes: 2
{ Number of separated
. roadways : 1
Position of the road:
level road
| Land cover type : rural
area
Figure 2 : The Segment A is selected
In white the CDB road segment, in black, the points after readjustment
on the center of the road.
Figure 3: The road following is processed (in black).
The CDB point (readjusted on the road center) is used as a road seed.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
Th
soi
col
tha
ge
ma
ho
dif
sh:
ori
(as
