Strategy: The strategy consists of two steps, namely the
detection of the buildings in the DEM and the reconstruction
of a parametric or prismatic geometric description of each
detected building.
The detection is achieved by computing an approximation of
the topographic surface using morphological filtering. The
difference of the approximation and the real surface gives re-
gions of potential buildings, which are filtered again using the
above mentioned thresholds concerning size and height. In
the second step - the reconstruction - the potential buildings
are analyzed according to height and shape. For each region
the principal axes are computed, along with width and length.
Depending on the height, either a prismatic or a parametric
house model is adjusted to the segments using the main axis
as ridge of the building. The method is described in detail in
[Weidner & Förstner 1995] and [Weidner 1995].
Results: The algorithm works fine on the data set where
the assumptions are clearly followed. All the 17 buildings
could be detected and reconstructed. Investigations on the
sampling size of the DEM show that the higher the grid size
the better the results are.
Figure 1: Detection and reconstruction of buildings in data
set flat
In the second data set (suburb) however, the buildings are not
that distinct both in height and separation from each other.
Therefore the algorithm fails partly to detect all the buildings
and is not able to reconstruct them correctly.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Figure 2: data set suburb: although many buildings are de-
tected, most of them could not be reconstructed properly
2.3 Norbert Haala
The approach is similar to the previous one in terms of the un-
derlying object model, but differs in the reconstruction phase.
Data Source: Stereo image pair and range data, data sets
flat and suburb
Object Model: The approach also bases on the generic
assumption, that buildings are usually higher then their sur-
rounding topographic surface. As a description parametric
building models are used, namely rectangular buildings with
either flat or symmetrically sloped roofs.
Prior Knowledge: The minimal size of the buildings is 10
m? and their minimal height is 3 m.
Strategy: This approach bases on a fusion of stereo-image
and range data. The strategy consists of two steps, namely
the detection of the buildings in the DEM and the reconstruc-
tion of a parametric geometric description of each detected
building in the stereo images.
The detection is achieved by computing regions of interest
in the DEM. To this end, regions being higher than their sur-
rounding are extracted. Regions of a certain height differ-
ence and size are investigated in the following reconstruction
step. The reconstruction is performed in the stereo image. In
both images straight lines are extracted and matched to form
3D-lines. The matching makes use of the approximate paral-
lax given by the DEM. In the final step a parametric building
model is approximated to this set of 3D-lines. The building
with the best fit in terms of minimal errors is chosen to rep-
resent its correct reconstruction. Another approach aims at a
segmentation of the DEM alone, namely extracting 3D-lines
directly from the range data set and fitting the parametric
model to these lines.
A detailed description can be found in [Haala 1995] and in
[Haala & Hahn 1995].
Results: The algorithm was tested both on the flat and
the suburb data set. In both cases, good results could be
achieved. In the data set flat all 17 buildings could be de-
tected and reconstructed, in the data set suburb 30 from 38
770
were
strate
only «
data,
metri
Figure
buildi
2.4
Data
Objec
jects,
projec
pects
regior
consti
relatic
Prior
>
>
>
Strate
step, \
to the
estim:
to the
given
Resul
This is
image
mode!
the ro
the se
be ext
For sc
a "fea
times
+ shac
Some
due tc
Will re
is due
exam
the he
Z para
