bel 132
Figure 6: Projected models
Figure 6 shows extracted building models (white lines) pro-
jected into the left image of the stereo pair. A qualitative
evaluation indicates that the orientations of the extracted
models fit to the image information. A rough comparison
of the extracted roof heights with manually measured points
indicate correspondence. The mean of the differences (abso-
lute value) is about 0.2 m for the ridges and 0.5 m for the
eaves. Problems occur for the paramters lenght and width,
although an overlay (Figure 5) of the DSM and extracted
ground plan information indicates a plausible fit. An expla-
nation of the effects may be that during DSM generation,
interest points are found at the borders of the roofs. Due to
low texture (c.f. label 113) or shadows (c.f. label 134), no
interest points are found close to the building for supporting
matching. Therefore, the regularization term within the re-
construction algorithm leads to interpolation between points
at the roofs’ borders and points on the ground, which are
more or less far away from the building, thus elongating the
buildings sytematically. Furthermore, the round offs at break-
lines contribute to such effects, although we try to take these
effects into consideration during the refined segmentation.
Fig. 3 (Reconstructed Polygon) displays the extracted poly-
gon superimposed on the original range data’, acquired
by airborne laser scanning. For the data set local MDL-
application leads to a reduction of the number of points from
98 to 36. The hypothesis about geometric relations between
edges of the polygon, which are introduced in the robust
estimation, put constraints onto the edges. A qualitative
evaluation shows little discrepancies, whereas the overall per-
formance seems to be acceptable. The discrepancies are on
one hand due to the sequence of analysis steps used here
(c.f. Section 4.2). On the other hand not all hypotheses
passing through the robust estimation are actually correct.
6 GIS DATA AND CHANGE DETECTION
In our approach to building extraction from DSM GIS or map
information can be incorporated as additional source of infor-
mation. GIS and maps mainly deliver 2D information about
buildings. The information about the third dimension in a
GIS is often related to the topographic surface, represented
! The range data of Hannover with a ground resolution of 2 m was sup-
plied by Dornier, Friedrichshafen.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
by a DEM, or this information may only be qualitativ, e.g. the
number of floors of a building, which can be used to derive
quantitative height information, if the mean height of floors
is known.
wr 295 — A
Hypotheses about changes
ESSEN
SS "v^
Difference
Figure 7: Use of DSM and GIS/map information
The DSM description of a scene has one advantage compared
to the GIS data, because it describes the actual scene which
might differ from the GIS due to changes. Therefore, GIS
data and DSM can be used for two purposes:
e The 2D information about buildings in a GIS or map —
depending on the scale — can be considered to be more
precise than the 2D information which can be extracted
from a DSM, and can therefore replace this informa-
tion. The DSM only serves as information source about
the third dimension for the buildings in the GIS.
e The DSM can be used to generate hypotheses about
changes in the scene.
A possible scenario is presented in Figure 7. The principle
idea of this scenario is to apply our approach without using
information from GIS and compare the results with the results
using this information. For this purpose the results are repre-
sented in DSM and the comparison consists of computing the
difference between these two results. Binarization and a com-
928
ses
abl
ick
dis
wh
Po
dis
veı
in
nes
of
str
rov
ou
inv
-— OR
T=
