ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002 
  
which are still connected to buildings. In a final stage of 
analysis, we try to eliminate such areas. By morphological 
opening using a square structural element, regions just 
connected by small bridges are separated. The resulting binary 
image is analyzed by a connected component analysis which 
results in a greater number of regions, and the terrain roughness 
criterion is evaluated again. Pixels being in regions now 
classified as containing vegetation are erased in the initial 
building label image. Thus, in vegetation areas originally 
connected to buildings, only the border pixels remain classified 
as "building pixels". Again, morphological opening helps to 
erase these border pixels. The resulting building label image 
only contains a small percentage of erroneously classified pixels 
in some backyards (Figure 5e). 
At a very coarse level of detail, a 3D city model can be derived 
by creating prismatic models from the boundary polygons of the 
building regions using the average building heights computed 
from the DSM. An example for such a city model with a height 
accuracy of about +5 m is shown in Figure 5f. 
5. GEOMETRICAL RECONSTRUCTION OF 
BUILDINGS 
5.1 Generation of initial 3D planar segments 
To start with model generation, initial 3D planar segments, their 
geometrical parameters, and their initial border polygons have 
to be found in the regions of interest. This is achieved by 
generating a “segment label image” defined in object space with 
an appropriate grid width. Each pixel of that image is assigned 
the label of the planar segment it belongs to. 
The framework for polymorphic feature extraction (Fuchs, 
1998) is applied for the generation of planar segments, too. Just 
as described in section 4, the framework is applied to the first 
derivatives of the DSM, this time using a small integration 
kernel of 3 x 3 pixels. Pixels classified as being homogeneous 
are surrounded by pixels having similar components of the 
normal vector, i.e., they are in a region containing co-planar 
points (Brunn and Weidner, 1997). The binary image of the 
homogeneous pixels is used for further processing (Figure 6a). 
By applying a connected component analysis to this binary 
image, planar patches should be detectable. However, due to 
classification errors, especially at the intersections of roof 
planes which are almost horizontal, the regions thus detected 
often turn out to be too large. Typically, this leads to L-shaped 
segments such as the one at the upper left corner of Figure 6a. 
In order to avoid these segmentation errors, an iterative strategy 
is applied for the generation of planar patches: 
1. The binary image of homogeneous pixels is 
morphologically opened using a square structural element 
before applying the connected component analysis. 
2. The geometric parameters of the planar patches thus 
detected are derived along with their r.m.s. errors from all 
points inside these patches. 
3. The patches with the best fit, i.e., those with r.m.s. errors 
better than a certain threshold (e.g., £10 cm) are considered 
to be seed regions for region growing. These seed regions 
are grown iteratively by adding neighboring pixels to a 
region if their distances from the original adjusting plane 
are below a certain threshold. In this way, the most relevant 
and best fitting planes are extracted from the DSM. 
4. The plane parameters are updated, and the pixels already 
being assigned to a planar patch are erased in the binary 
image. The connected component analysis is repeated, and 
the parameters of the new planar patches are evaluated. 
Steps 1 to 4 are repeated with a decreasing size of the structural 
element for morphological opening. Thus, smaller and smaller 
initial regions are found, and by only allowing well-fitting 
planes to grow, it is possible to split the regions corresponding 
to more than one roof plane, because the r.m.s. error of the 
planar fit is a good indicator for the occurrence of such 
situations. Figure 6b shows the planar patches extracted in one 
of the building regions from Figure 5e. 
  
Figure 6a. Pixels classified as “homogeneous” (white) for one 
of the building regions in Figure Se. 
  
Figure 6b. Planar patches obtained by iteratively applying 
region growing. 
  
Figure 6d. Final segment label image for one of the building 
regions in Figure Se. 
A further analysis has to detect planes which cover too small an 
area for resulting in pixels classified as being homogeneous. We 
search for regions not being consistent with the planar regions 
detected so far (Figure 6¢). The borders of the buildings are 
typically found in that process, which is caused by laser points 
on the walls and by the effects of grid interpolation. Again, we 
get rid of these points by a morphological opening operation 
using a 3 x 3 square structural element, and a connected 
A - 299