then computing the final coordinates in the appropriate
cartographic projection.
The method we use to match roads from both images is based on
a local correlation which takes into account some contextual
knowledge about the detected road : the last measured disparity,
the road orientation and the road width are used to predict the
image area which has to be correlated :
e road orientation is used to compute the correlation
coefficient on elongated rectangular areas.
e road width is used to give more important weight to the road
edges.
e last measured disparity is used to limit the search area
through a maximal slope threshold.
The coordinates transformation from the image to the
cartographic projection is not a problem since we know the
image absolute orientation.
2.4. Implementation in SAPHIR Géo
This semi-automatic road extraction method has been
implemented in SAPHIR Géo by SYSECA. The interactive part
of the capture process needs a well adapted environment to be
efficient :
e user-friendly human-machine interface, using a large range
of input devices (mouse, handwheel, tracker ball) and
multiple windows.
e 3-D display of the result to have a good mean of control,
using the DPW® software of Helava.
e optimized hardware and software to respect the real-time
constraint : real-time display of large volumes of information
in both raster and vector form, real-time management of
vector data in a GIS environment.
This interactive part of the data capture process is the only one
which requires such an environment. If we decide to use as input
of this step an automatically pre-computed graph of the road
network, this graph may have been computed on any other
environment, with no particular specifications.
3. THE AUTOMATIC INITIALIZATION
3.1. Using the image only
The first solution we propose to compute a reliable incomplete
graph of the road network is based on the traditional steps
described in the literature (Aviad, 1988 ; Gruen, 1995) :
e road finding.
e road tracking.
e road linking.
The first step is supposed to extract the most visible road
portions using local properties. The second step propagates the
road presence hypothesis from the previously detected road
seeds. The last step introduces higher level constraints to re-
construct the road network as a topological graph with a few
interruptions.
Our approach may be divided into five steps (Figure 3). Each of
these steps uses very hard thresholds since we are looking for a
reliable result, even if this result is largely incomplete.
16
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
road seeds filtering
road following
elevations computation
i
road portions filtering
i
road network graph
left
image
right
image
Figure 3 : the automatic processing using the image only.
a) road seeds detection : the road seeds detection is based on the
shape analysis of the regions coming from a region-based
segmentation (Ruskoné, 1994).
b) road seeds filtering : the shape criterion used before is too
local and we have to apply other criteria to eliminate many false
detections. The criteria we use as a confirmation of the road
presence are :
e the contours around a road are geometrically organized : the
entropy of the gradient directions around the road seeds has
to be rather low.
e a road seed can not be a local maximum of elevation : using
both images, we compute locally the elevations and filter the
road seeds using a top-hat function.
c) road following : this step is based on the road following
algorithm described in the previous part (road detection, parallel
edges extraction, interpolation using splines).
d) elevations computation : the 3-D coordinates of the detected
roads can be computed using the same criteria as in the semi-
automatic approach (correlation using the road orientation and
the road width). As we work in this case with large road
portions, it is possible to optimize these criteria along the whole
object, using an additive constraint of regularity of the road
slopes. The model we use looks like the active contours model
with the opposition between internal constraints (minimizing
altimetric curvature) and external constraints (maximizing the
correlation between both images).
e) road portions filtering : this step has to mix the different
measurements which have been made along the whole
processing to take the final decision. A supervised learning has
been made for different kinds of objects (not only roads) in
different kinds of contexts and it is possible to classify
(Ruskoné, 1995) each road hypothesis and to compute a
probability to belong to the road class. Only very high
probabilities are kept.
3.2. Using external data
The other solution we propose to automatically generate a
reliable incomplete road network graph is to use external data as
input. In our case, these external data would be provided by a
cartographic database, which is already available and have a
geometric accuracy in the range of 10 meters. The external data
could also be provided by the automatic digitization of paper
maps.
POSE
The problem
the approxim
the network.
We propose a
left &
right
images
Figure
a) crossroads
processing to
is based on a
branches.
b) road porti
adjusted and 1
problem can
have to find
The computat
contours mod
images and an
This step can
before.
4.
It is now imp
automatic ext
processing. 1
consists in d
according to
parameters we
4.1. The indic
About the in
geometric qu:
database spec
saved with suc
e geometric
automatic
the geome
e succes rate
of the ini
number of
the numbe
e capture tii
important |
time with
rate of our
e User's rem
remarks 2
automatic
accepted a
use it (Mcl
