ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision“, Graz, 2002 
  
main steps: the segmentation and the detection of buildings. A 
segmentation software segments the height data. The vectorised 
object segments as well as the digital map are the inputs for a 
GIS software. The buildings are detected by combining the map 
information and the segment attributes. 
2 STUDY AREA AND DATA SET 
Two study areas, each with 7km? in the region Gruyères, 
Switzerland were chosen — a rural mountainous area and a flat 
urban region. The rural study area mainly consists of residential 
areas with mainly single-family houses in a typical regional 
style. For the urban study area, the city Bulle was picked out. 
Bulle contains beside single-family and multi-family residential 
units also commercial and light industry areas. 
The laser scanner data is first pulse data with a mean point 
distance of 1.2m. It was acquired in early 2001. The scanned 
topographic map was a small-scale map (1:25,000) from 1995. 
Due to generalisation effects, planimetric discrepancies of up to 
12 metres can occur between the map and the laser scanner data 
set. An aerial (false) ortho photo, from 1998, was used to check 
the results. 
In the beginning of the study, some statistical data was 
generated to confirm the main idea explained above. The rural 
study area contains 820 buildings derived from the DSM. In the 
pixel map, about 93% of the buildings are recorded. Due to 
generalisation effects, 3% of the houses are insufficiently 
positioned, that means that the centre of the building in the 
pixel map is more than four metres apart from the building in 
the laser scanner data. This needs to be considered in the 
analysis. The urban study area had 1630 houses in spring 2001 
of which 9% are not recorded in the map and 3% are 
insufficiently positioned. 
3 SEGMENTATION OF OBJECTS AND THEIR 
FEATURES 
3.1 Data Preparation 
Before the segmentation could be performed, data editing and 
formatting had to be done. The laser scanner data was 
interpolated into a regular grid and saved as an image with 1m? 
raster cells in order to be processed with raster oriented image 
processing tools. The pixel map required the most editing. The 
situation information such as houses, streets, or lettering was 
extracted. A morphological filter was applied, to mainly reduce 
this information for houses. The procedure consisted of an 
opening filter followed by a dilating filter. Thus, the extension 
of the houses is shortened, but this will not effect the analysis. 
Some parts of the map text are still in existence and need to be 
excluded from the analysis. 
3.2 Segmentation 
Most methods applied for the segmentation of laser scanner data 
interpolated to a regular grid, were purely pixel-based methods 
(e.g. Oude Eleberink and Maas 2000). The goal of the approach 
shown here was to extend these methods to region-based 
methods in order to improve the quality of the segmentation. 
The presently available region-based segmentation software 
eCognition 2.1, usually used with image data e.g. in (Neubert 
and Meinel 2001), was utilised. Image analysis with eCognition 
is based upon contiguous, homogeneous image regions, which 
are generated by an initial image segmentation. The image 
content is represented as a network of image objects. The 
segmentation algorithm of eCognition is based on the so-called 
“bottom up region-merging” technique. Thus, starting with the 
pixel level, each pixel is joined with its most similar neighbour 
pixel. In a number of steps, these two-pixel-objects are enlarged 
to bigger pixel groups until a certain heterogeneity value (scale 
parameter) is reached. These pixel groups (segments) are 
optimised considering the homogeneity criterions color and 
shape. With these three parameters (scale, color and shape) the 
user can influence the segmentation process. (Definiens 
Imaging 2001) 
Referring to laser scanner data, it is supposed that objects, 
having a different height in comparison to their surroundings, 
are assigned into one segment and the surroundings into 
another. 
The segmentation performed with eCognition requires some 
experience. As the scale parameter, which is a measure of 
heterogeneity, determines the size of segments, the user has to 
check carefully whether the segmentation was successful. The 
correct scale parameter can only be found empirically. The 
homogeneity criterions are easy to set, given that the height 
information is the only relevant one. Hence, the color is high 
weighted and the shape is neglected. Weighting the shape 
parameter would result in segments of similar shape and area. 
However, only buildings are supposed to be compact and 
rectangular, not their environment or vegetation. The shape of 
the segments is used in the later classification. Thus, weighting 
the shape parameter would complicate the classification. 
    
igure 3-1. Example of laser scanner data ( 1m raster) 
The experience shows that segmenting pure laser scanner data 
(Figure 3-1) does not always produce the desired result. The 
reason is the smooth transitions of the objects to their 
neighbours such as houses and trees that stand close together. 
Therefore, attempts were made to modify the laser scanner data 
to improve the segmentation result. No pleasing solution was 
found. To comprise the ortho photo and / or the pixel map 
directly in the segmentation was not valuable as planimetric 
discrepancies, mentioned in chapter 2, pervade the data sets. 
The segmentation statistic presented in Table 3-1 indicates that 
up to 85% of the buildings were segmented according to the 
features form. Passable segmentations occurred when trees 
A- 170