ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision‘, Graz, 2002 
  
   
  
    
D, LEECEEREBEREEEHEREEHHEREG 
Figure 4. Final DTM (0.5 m; here: 1.0 m for better readability) 
  
4. BUILDING DETECTION 
As described in section 3, two digital elevation models are 
derived by interpolation: a DTM is computed from the points 
classified as “terrain points” with a high degree of smoothing, 
whereas a DSM is computed from all points without smoothing 
(Figure 5a ). An initial building mask is created by thresholding 
the height differences between the DSM and the DTM (e.g., by 
Ahmin=3.5m). This initial building mask still contains areas 
covered by vegetation, and some individual building blocks are 
not correctly separated (Figure 5b). A morphological opening 
filter using a small (e.g., 5 x 5) square structural element is 
applied to the initial building mask in order to erase small 
      
* 
as = x à = *s 
. 
m € ki: xf t ! # a^ - 
pw et. i. > den ay * An ; 3 
+ yd WR Reo ^ * " » 
fs wu maf 
    
elongated objects such as fences and to separate regions just 
bridged by a thin line of pixels. A connected component 
analysis of the resulting image is applied to obtain the initial 
building regions. At this instance, regions smaller than a 
minimum area (e.g., 40 m?) and regions at the border of the 
DSM are discarded (Figure 5c). 
Some of the remaining regions in Figure 5c still correspond to 
groups of trees. These regions can be eliminated by evaluating a 
"terrain roughness" criterion derived by an analysis of the 
second derivatives of the DSM. In (Fuchs, 1998), a method for 
polymorphic feature extraction is described which aims at a 
classification of texture as being homogeneous, linear, or point- 
like, by an analysis of the first derivatives of a digital image. 
This method is applied to the first derivatives of the DSM using 
a large (e.g., 9 x 9) integration kernel. For each initial building 
region, the number of “point-like” pixels is counted. Regions 
containing more than 50% of pixels classified as being “point- 
like” (thus, pixels being in a neighborhood of great, but 
anisotropic variations of the surface normals) are very likely to 
contain vegetation rather than buildings, and they are 
eliminated. Figure 5d shows the results of texture classification. 
Note the obvious co-incidence of clusters of “point-like” pixels 
displayed in black and vegetation areas such as those in the 
Belvedere gardens on the left margin of the test site. 
The terrain roughness criterion is very efficient in classifying 
isolated vegetation regions, but it cannot find vegetation areas 
  
Figure 5. Building detection in a test site in the City of Vienna. Original resolution: 0.1 m (in-flight) by 1.0 m (cross-flight). 
a) DSM; grid width: 0.5 x 0.5 m?; extent: 410 x 435 m. b) Initial building mask (height threshold Ah,,,,73.5m). 
c) Initial building label image before evaluating terrain roughness. d) Results of texture classification. Integration kernel: 9 x 9 
pixels. White: homogeneous; gray: linear; black: point-like. e) Final building label image. 14 building regions have been detected. 
f) VRML visualization of prismatic models created from the boundary polygons of the building regions from Figure 5e. 
A - 298