ES
3
/
)
yomson. fr
) the photogrammetric data
ability delay. Since digital
semi-automatic tools in the
he road network automatic
interpretation of the image
is, this paper concerns the
efficiency of such kind of
1s in a production context.
cernant l'automatisation de
sponibilité des données sur
raînes de production, il est
onnels de saisie. L'un des
availlons parallélement sur
sur des approches semi-
sente différentes solutions
ugmenter l’efficacité de ce
solutions dans un contexte
acquired in the production
platform which has been
is developed by SYSECA.
l-accuracy softcopy image
rates a comprehensive
1g package with an object-
eo-referenced objects.
h laboratory of IGN which
topographic database data
t advanced project in our
. automatic extraction. We
ry ambitious approaches,
tation of the image and on
or a short term efficiency.
| been presented in the
made objects from aerial
n to work both accurately
1ages and objects (Hsieh,
roblem of the automatic
rpretation and Remote
mages pour la Stéréo-
nna 1996
interpretation of aerial images, many authors present the semi-
automatic approaches as the only ones able to bring a significant
contribution in a production context in the years to come
(Heipke, 1994 ; Gruen, 1995).
We proposed (Airault, 1994a ; Airault, 1994b). and evaluated
(Airault, 1995) a semi-automatic solution in former
publications. The evaluations we performed have shown that it
seems to be possible to save time with our semi-automatic
solution if the road following algorithm is used in a good way :
only when the road in the image corresponds well to the road
model used in the algorithm. The efficiency of the semi-
automatic technique will much more depend on its reliability
(predictable behavior, low sensitivity to small changes in the
parameters, preference to no result than to a wrong one...) than
on its exhaustiveness (Jamet, 1995b). Nevertheless, even if the
results of our evaluations were very optimistic according to the
automatic extraction geometric quality, these results were rather
modest according to the efficiency in terms of productivity
enhancement (from 10% to 40% of saved time according to the
landscape characteristics). Independently of the robustness and
of the efficiency of the detection algorithm itself, a semi-
automatic work session takes time from the user, at least the
time taken by all the moves into the image and the time taken by
the user's decisions.
1.3. How to increase the automation level ?
To enhance significantly the productivity, we have to increase
the automation level.
One solution consists in keeping a semi-automatic approach but
which would be initialized by a pre-computed incomplete road
network graph (Figure 1) :
e fully-automatic extraction of a little part of the network.
e interactive completion using a road following algorithm.
| fully-automatic initialization |
road network graph
Y
| semi-automatic completion |
ri ess | ES E s -: i Semi-automatic :
control : : corrections road following
te cueene ee uere cac 00000000 4
completed road network graph
Figure 1 : semi-automatic completion of a pre-computed graph.
If fully-automatic systems are not ready to extract the whole
road network with a good reliability, it seems to be possible
from now to extract automatically the "easiest" ones with a very
low error rate, applying very hard constraints. A so extracted
part of the road network can then be presented to the user for a
visual control and a semi-automatic completion.
2. THE SEMI-AUTOMATIC SOLUTION
The semi-automatic approach we propose relies on several
algorithms used sequentially on the same platform.
The main originality of the approach is to separate the detection
step and the geometric adjustment step (Airault, 1994a). The
complete processing can be divided in five distinct (and
successive) steps (Figure 2) :
col prima (t)
lig
par
rw road detection
pt. ima (t1)
l
| parallel edges extraction |
pt_ima
pt_ima (t+1)
J
| geometric adjustment |
pt_ima (t+1)
l
| parallax estimation |
pt_map y pt_ima (t+1)
: J
| cartographic projection |
i
pt_map(t+1)
Figure 2 : Modular decomposition of the semi-automatic
road following algorithm.
The data structure pt ima corresponds to the points
characteristics in the image. It contains, in addition to the
location in the image, the disparity (or parallax) value and the
measured road width. The data structure pt map corresponds to
the cartographic coordinates of the road points.
2.1. The detection step
The detection step corresponds to the identification of the image
area as a road object. This step is based on the local
optimization. of a cost function The used cost function takes
mainly into account the local homogeneity and the anisotropic
aspect of the homogeneity of roads. It takes as input one point
and one direction which can be given by the user (as an
initialization) or which can be the previous result of this iterative
processing. It gives as output the new possible location on the
road and several indicators which allow to decide if the plotting
is still on the road.
2.2. The geometric adjustment step
The geometric adjustment step is based on a local detection of
parallel edges along the road. The whole step is designed to
compute a precise location of the road axis, using both the
detected edges and an a priori geometric model of roads (as
cubic splines). The road edges are extracted as often as possible
(they do not always clearly appear in the image) and the road
model is used to interpolate the road position between these well
positioned points.
2.3. Computation of the 3-D geometry
The previous steps are using only one image to detect the road.
The goal of the next ones is to compute the 3-D coordinates of
the road, trying to match the detected road with the
corresponding object in the other image of the stereo pair and
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
