racted from
disparity
y is known
sroads and
map of dis-
rformed on
ach point is
angle repre-
| associated
coordinates
nsider com-
ier patches
e].
e suburb of
of construc-
and shapes.
on figure 1,
; part of the
fter the ma-
ad sections.
en the map
ated with a
ally selected
rk extracted
osed on the
as described
out the va-
is presented
errors given
300
|
number of points
200
|
100
|
|
-10 -6 o 5 10
error (meters)
Fig. 8- Error histogram between
ground truth and generated DTM.
by wrong disparity values at crossroads. These errors
generate pitches or holes in the DTM.
The road section validation process confirmed the
disparity for 330 crossroads after the first iteration
and for 348 crossroads after the second iteration. Fi-
gure 10 presents the final DTM on which red features
of the map (main roads and contour lines) have been
superimposed. One can notice that the terrain model
fits very well the contour lines, and that the errors
previously seen on figure 7 have been removed.
4.2 Performance analysis
Evaluation of these results has been made against
a ground truth given on grid of 25m x 25m resolution.
The left image in which referential the DTM is cal-
culated has been calibrated with these ground truth
data. Each point in the ground truth is given by its co-
ordinates (.X, Y, Z) in a cartographic referential. The
calibration gives the coordinates (u,v) of the corres-
ponding pixel in the left aerial image. It is then pos-
sible to compare the ground truth elevation Z of the
point with the elevation Zg;,,, given by the calibra-
tion of the pixel disparity. The error at a point is then
defined as the value of the difference between these
two elevations:
€or Z dispa —-
This error has been calculated for the 5489 common
points of the ground truth and the generated DTM.
The mean of the error absolute value is equal to 2.10m.
Figure 8 presents the error distribution for the scene.
As it could be expected this distribution presents a
peak centered on zero, but one can notice that other
significant peaks are present for positive error values.
This indicates that our elevation estimation method
tends to over-estimate the altitude of points.
At this point of our research further considerations
should be undertaken in order to identify these errors
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
9 ON OE 6 06 9
Fig. 9 - Error distribution in the scene,
dark points indicate large errors.
more precisely. Figure 9 presents the error distribu-
tion in the scene. We can first notice that errors occur
on the sides of the scene where crossroad disparity is
validated with less road sections (see 3.2). But largest
errors appear on the highest region of the scene where
less crossroads are present, because of a building-free
area. Future developments for our DTM generation
system consist in the analysis of the white areas of
the map in order to compensate the lack of crossroads
in these regions.
5 Conclusion
In this paper we describe a new method for automa-
tic. digital terrain model generation using information
provided by scanned maps and stereo pairs of aerial
images. Scanned maps provide useful information on
the location where the terrain could be directly seen on
aerial images, mainly roads, crossroads and building-
free areas. We demonstrated theses ideas with complex
urban scenes presenting a large variety of construction
density and buildings of various sizes and shapes.
Further developments are under progress in order
to improve the quality of the DTM:
— use of Bernstein-Bézier patches in order to ob-
tain a G! continuous surfaces,
— finer disparity analysis in building-free areas in
order to compensate the lack of crossroads.
D'TM generation is the first part of a more global
system dedicated to the construction of 3D cartogra-
phic databases using simultaneously aerial images and
scanned maps. Further developments are dealing with
the generation of building 3D models, for which scan-
ned maps are also a rich source of information on the
presence, shape, size and localization of buildings in a
dense urban scene.
