ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002 
  
  
Figure 3-5. Example of aspect based roof segments 
4 DETECTION OF BUILDINGS 
In a next processing step, a GIS package is used to detect 
buildings in the segmented regions. The segments are analysed 
by their position, shape and attributes. Rules have to be found 
and complex relationships have to be recognised. To find as 
many houses as possible, several methods have to be combined. 
First, the houses recorded on the map are located in the roof- 
segment-file. Their centre coordinates are saved in a look-up 
table. Secondly, the house-segment-file is classified and centre 
coordinates of possible houses are compared to those in the 
table. Those not in the table will be added. Later the single 
segments are analysed whether they hold a building or not. 
4.1 Detection by Map 
For this part of the analysis, the modified and vectorised pixel 
map is utilised. Since the filter routine did not remove all no- 
  
house information in the rural part, these superfluous segments 
have to be eliminated or at least reduced. Analysing the 
standard deviation values of the surrounding laser scanner data 
does minimize them. Those segments in the roof-segment-file, 
which are in a distance of four metres to the map segments, are 
selected. An ID is assigned to the roof segments according to 
their position to the map segments. After dissolving the roof 
segments, the following parameters are chosen to eliminate 
polygons that are located in field or forest areas and cannot 
represent houses: the average standard deviation of the laser 
scanner data per segment and the average of the slope values in 
a segment. The value of the average standard deviation of the 
laser scanner data is low for field and high for forest areas. 
Areas of small bushes have a mean average standard deviation 
for laser scanner data — a low average slope value will reveal 
them. In this step of the analysis, the laplacian values are not 
valuable, because of the size of the dissolved segments and the 
spatial resolution of the data set at hand. 
The selected no-house polygons are excluded from the map- 
segment-file. About 48% of the no-houses map polygons are 
eliminated and all house segments are retained, which is most 
important. Table 4-1 presents the statistics of the pixel map 
derived polygons. 
The procedure was not necessary for the urban study area, as the 
morphologic filter worked adequately. 
  
  
Number of Number of Other 
buildings misplaced objects 
buildings 
Rural area 768 (93%) 26 (3%) 455 
Urban area 1434 (91%) 41 (3%) 193 
  
Table 4-1. Modified pixel map statistic 
The remaining map segments are fed into in a routine, which 
selects all polygons in the roof-segment-file. To make sure that 
the planimetric discrepancies between map and laser scanner 
data do not prevent the selection of all roof segments of a 
building, all segments in the roof-segment-file that are within a 
certain distance of the map segments are selected. For normal 
houses — single-family houses — a distance of 2.5 metre is 
sufficient, whereas buildings larger than 300m? require a 4m 
buffer zone. The centre coordinates of the selected roof 
segments per map segment are stored. 
The segments incorrectly marked as houses, will be thrown out 
of the analysis by the later house modelling we are still 
developing. 
4.2 Detecting by Attributes 
Houses not detected via the pixel map are to be found by 
analysing the house-segment-file. The procedure is comparable 
to a classification. The house-segment-file supplies valuable 
attributes for the classification such as mean value and standard 
deviation of the input laser scanner data, mean value and 
standard deviation of the aspect, slope, and laplacian image, and 
shape attributes such as area, border length, and main direction. 
In accordance with the attributes, each segment is signed as 
house-containing or no-house-containing. 
Within the classification, the following assumptions and rules 
are made for rural regions: 
— Buildings are not to find within forest or water areas 
— Building segments are smaller than 6000m? 
A- 172