In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C. Tournaire O. (Eds). 1APRS. Vol. XXXVIII. Part 3A - Saint-Mandé, France. September 1-3. 2010
the best planar fit (indicated by s n ) is accepted as a seed region
for region growing. Points not yet assigned to any other plane
are added to the seed region if they are found to be coincident to
the plane based on a statistical test. The resulting point set is
analysed whether it corresponds to a connected segment in
object space. If it is split into several segments, the segment
containing the largest number of the original N WI nearest
neighbours is maintained. The parameters of the adjusting plane
are determined from all points contained in the segment, and the
planar segment is added to the combined label image. This
procedure is repeated until no new plane can be found.
Figure 4. Left: the left projected label image from Fig. 1
without the segment from Fig. 3 and without ground
segments. Centre: multi-image segmentation. Right:
results after merging co-planar segments.
By erasing the segments generated by region growing in the
projected label images, image segments corresponding to
several roof planes may be separated. Thus it makes sense to
repeat multi-image segmentation, this time using NP min = 3 so
that also very small segments can be detected. Finally, we try to
close small gaps in the segmentation results by checking
whether there are points that have not yet been assigned to any
segment, but are coincident to one of the planar segments in
their vicinity. These points are assigned to the nearest segment
in terms of the point's distance from the plane. Again, co-planar
segments are merged. At this stage, the segmentation results
may contain segments that correspond to roofs of neighbouring
buildings if they are close together, or they may even contain
segments on vegetation or other high objects in the vicinity of a
building. We eliminate planes having an overlap smaller than
O min of the segment’s area with the region of interest defined by
the approximation. No further classification is earned out.
The left part of Fig. 5 shows the results of the region growing
process for the building in Fig. 1. Three new planar segments
were added by the region growing process, two corresponding
to the dormers and one that merges two roof planes where they
intersect at an obtuse angle. After that, the second multiple
image segmentation process can find a few more roof planes,
because they have been cut off from the large image segments
in the shadow areas (central part of Fig. 5). The rightmost part
in Fig. 5 shows the final segmentation results after filling small
gaps and merging co-planar segments.
3. RESULTS AND DISCUSSION
In order to test the method described in this paper, it was used to
detect the roof planes of seven buildings in the historic centre of
Vaihingen, using the data described in Section 1.3. The
approximate building outlines were generated by the building
detection algorithm described in (Rottensteiner et al., 2004).
The planimetrie accuracy of these outlines is about ±2 m. The
buildings in this area show medium to high complexity. The
number of images a building was visible in varied between six
and nine, because the buildings are situated in the overlapping
area of three strips. Tab. 1 shows the parameter settings used in
our experiments. The results of our method, one of the images
used for achieving those results, and a reference segmentation
based on photogrammetric plotting are shown in Fig. 6.
Figure 5. Left: results after adding new segments based on the
point cloud to the results in Fig. 4. Centre: results
after the second segmentation. Right: results after
merging co-planar segments and filling small gaps.
Figure 6. Left: one of the aerial images. Centre: segmentation
results. Right: reference. FP: False positives. TN: True
negatives (missed planes). T: Topological errors.
Fig. 6 shows that multi-image segmentation does a good job in
separating the individual roof planes. This is confirmed by a
