vel 
s to 
> of 
'eas 
of 
| on 
put 
this 
and 
tric 
n is 
  
ved 
leir 
ees 
the 
on, 
we 
the 
Petra Zimmermann 
  
3.6 Shadow 
Shadow causes problems, e.g., in line following and segmentation. Shadow is no cue as the previous explained cues; it 
has to be recognised through combining several basic cues. We use the HSV colour space colour segmentation (Figure 
14), edge information and information from the DSM and information about the date and time to derive areas occluded 
by shadow. At the moment our results have to be improved, because shaded parts of roofs and trees are still 
misclassified due to their saturation value. 
4 REALIZATION OF BUILDING DETECTION 
For building detection we extract coarse regions, store them with their bounding box, image primitives, boundary and 
attributes derived from different cues (Figure 16, 17). The buildings are detected through fusion of multiple cues as 
colour, edges and texture derived in different colour spaces and DSM data, see Table 1. In a first step we use edge 
detectors, texture filters and classification algorithms to derive independent low-level image primitives from grey-level 
and colour images. Edge detection gives us edges with the brightness or colour information in the neighbourhood, 
colour and texture filtering provides homogeneous textured or coloured regions. 
A DSM is used to identify objects with 
elevation above ground, derive their coarse 
shape, height, values of slope and the main 
axes. For the later fusion of these primitives 
we derive the quality of the primitives 
dependent on the cue and the algorithm and 
store it with the primitive data. A 
knowledge ^ base defines expected 
characteristics of man-made objects and 
buildings e.g. minimal size, homogeneous 
texture, minimal/maximal height, edge 
shape and length and supports the separation 
of man-made objects with elevation from ManMadeObject 
natural objects. Region attributes e.g. 7 
texture and colour are incorporated 
implicitly in classification algorithms to 
fuse these attributes. Region segmentation is 
done with different colour channels, 
elevation and texture information with 
thresholds for merging and weights for 
different channels. Region-based and 2D (c 
edge information is fused in a region- RoofPlane RoofBoundary 
refinement step providing homogeneous Figure 16: Process of extraction of features for building detection 
regions with exact boundaries e.g. roof 
planes, outlines and ridgelines. Depending on the thresholds, roof details could be computed. For later reconstruction 
steps objects are stored as polygons with attributes about quality, channels, thresholds and algorithms which were used. 
As a result of building detection the system knows about the location of buildings, relevant edges and relevant regions 
with their attributes. The resulting attributed image primitives can be used directly to derive 2D and 3D information and 
give constraints for the further reconstruction step. A control mechanism provides rules (preferences) to weight the 
primitives in fusion. If there is e.g. a region with high elevation detected in the DSM and in the same region were 
extended edges detected, this could be an indicator for a building and the system will look there for homogeneous 
regions. But if it detects extended edges but no elevation, the region might be a road. 
   
    
      
      
d 
RoofRidge 
4.1 Putting it all together 
The size of the assigned attributes is chosen with respect to their relevance for building extraction through all 
algorithms and normalised to values between 0 and 1. Different cues get different weights. According to Table 1 and 
Figure 16 we compute for each region dynamically whether a region could belong to a building or not through a basic 
Quasi-Bayes Network for each region, provided with a set of initial probabilities computed by the regions attributes. An 
"BuildingRecognitionA gent" gets initial preferences, a table of values for states (-values) of single attributes (Table 2, 
e.g. “Segment is inside/outside DSM blob” = 1/-1) and decides hierarchically through the set of attributes whether he 
accepts a region as “BuildingRegion” or as “Non-BuildingRegion”. In Figure 17, e.g. segment B fulfils in each step the 
preferences of the agent also for segment A; segment C and D will be rejected. A GUI (graphical user interface) allows 
the user to visualise the results, check single results and their attributes. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 1069