International Archives of the Photogrammetry, Remote Sensing and Spatial Information
Sciences, Vol XXXV, Part B3. Istanbul 2004
Figure 3: Results of ground level extraction for a sub-area of ‘the airborne laser DEM of the Amiens region. (a) Part of an aerial image
showing the sub-area of the Amiens region. (b) Part of the original DEM with altitude coloring. (c) The DEM part after smoothing
and isolated point removal. (d) The result of automatic segmentation. (e) The automatically extracted ground level. (f) The extracted
ground level with texture mapping for visual verification.
be a 3D modeling and editing environment which allows easy and
simultaneous visualization and manipulation of 3D data in con-
nection with (one or more) corresponding images of the scene,
hereby automatically maintaining and updating structural rela-
tionships between the image features in the different views and
with the 3D data. The different coloring modes used in these
examples are basic features of ReconLab. But ReconLab also al-
lows to map the image texture to the 3D information. Figure 3 (f)
shows the same ground segment as in Figure 3 (e), but with the
image texture mapped onto it. Zooming in to the red spot location
proves that it indeed corresponds to a built up area, as expected.
The editing capabilities of ReconLab allow to delineate and re-
move the erroneous area from the DEM segment with just a few
mouse clicks. In this way, a fast and easy high-level user inter-
action between the automated components of the algorithm (i.e.
segmentation and DTM surface fitting respectively) guarantees
a qualitatively optimal surface model for the terrain. But even
without this user interaction the DTM fitting algorithm does not
suffer much from the possibly remaining errors in the selection
of ground points, as is demonstrated next.
The segment corresponding to the ground level is used to initial-
ize the DTM estimation algorithm described in section 3. Fig-
ure 4 shows the DTM surface model that was computed from the
DEM points contained in the ground level segment of Figure 3 (f).
For visualization purposes, relative altitude coloring was applied
to the model. Therefore, the color pattern in Figure 4 is not the
same as that in Figure 3 (e).
The usefulness of prior ground level extraction to DTM estima-
tion has been tested on the other datasets as well. In particular,
the algorithm was applied to each dataset twice: Once with the
full DEM used for intialization, and once with starting from the
DEM points contained in the ground level segment only. In the
latter case, the initial DTM is a much better approximation of
the real terrain, thus allowing to use a smaller maximum value
of c and to have better and faster convergence of the algorithm.
Apart from this, the algorithm was in each case applied with both
N = 1and N = 2 as the order of the DTM model. Execution
times and number of iterations were recorded and compared for
each test. Moreover, all tests are performed on an Intel Pentium
[V processor running at 1.6 GH z, and a base 100 of normalized
execution time was used for the N = 1 DTM with initial data.
The results are summarized in Table 1. Observe that convergence
of the algorithm is improved by at least a factor 3 when initializ-
ing the DTM model with only the DEM points contained in the
ground level segment provided by ReconLab, and without loss of
accuracy. In fact, the difference between the DTM parameters Ó
is less than 2 96 when obtained with or without initial segmenta-
tion. This also demonstrates the power of the DTM estimation
algorithm in eliminating the above ground points during iteration
when starting from the full DEM.
5 CONCLUSIONS
A user-friendly software system for semi-automatic real time fil-
tering and segmentation of Digital Elevation Models is presented.
The system is capable of extracting the ground level points from a
dense DEM of complex urban areas, which show large variability
in landscape and in terrain slope and altitude. The required user
interactions are all high-level and mainly involve the supervision
of the process. The quality of the extracted ground surface points
is demonstrated by the fact that estimating a parametric DTM sur-
face model from these points requires a computation time which
is at least 3 times faster than without this preprocessing; and,
there is no loss in accuracy of the resulting DTM. These observa-
tions were corroborated by test on a synthetically generated DEM
and on real world DEMs obtained from airborne laser scanning
as well as by stereo correspondence from imagery.
REFERENCES
Baillard, C., Dissard, O. , Jamet, O., and Maitre, H., 1997. Ex
traction and characterization of above-ground areas in a per
International A
with in
N
Gar md
100
182 it.
H
with in
NE
Cmax — 4
100
54 it.
with in
N
Gar = >
100
23 it.
Table 1: Comp
tions for the D
level segmentat
X
e
Figure 4: The D
DEM points inc
Relative colori
urban context, i
1997. Mapping
tures from. Imc
puter Gesellsch
pp. 159—174.
Baltsavias, E.P.
Extraction of M
(III), Balkema |
Belli, T., Cord,
for stereo imag
Colour in Grap
Etienne, France
Belli, T., Cord,
tion and modeli
2001), pp. 125-
Collins, R.T., I
1995. Automat
images, in (Grü
Cord, M., Jorda:
Ing structure rec
puter Vision anc
