ISPRS Commission III, Vol.34, Part 3A »Photogrammetric Computer Vision“, Graz, 2002 
— The length/width rate is smaller than a certain value, 
determined by the longest building 
— The standard deviation of laplacian values are within a 
certain range for buildings 
Segments exceeding 6000m° usually represent field or branched 
street areas. The length/width criterion excludes long narrow 
segments, as they are to be found on grassland and streets. The 
standard deviation of the laplacian values of the segments is 
small for flat areas and high for undulating regions. To apply 
the laplacian value of a segment as a classification criterion 
works only with a good segmentation result. The building has to 
be entirely within a segment and the segment should not be too 
long-ranged around the building. 
Urban areas subject to different conditions. Thus, the 
classification parameters have to be altered to: 
— Buildings are not to be found within forest or water 
areas 
— Buildings do not have higher neighbour objects than 
buildings itself, except they are situated on a slope — 
thus the relative border to higher neighbours in the 
laser scanner data segments should not exceed a 
certain value 
— Building segments are smaller than 6000m? 
— The length/width rate is smaller than a certain value, 
determined by the longest building 
— The standard deviation of laplacian values are within a 
certain range for buildings 
Similar conditions can be suggested for flat, rural regions. In 
general, the detection algorithm is more complex for urban 
areas since the variation of house types is bigger — big flat 
roofed buildings of light industry units and small steep roofed 
houses in residential areas. Specific shape attributes such as 
number of arcs were not used since the shape of house- 
containing segments is not consistent (e.g. to buildings attached 
trees). 
The whole processing chain was optimised to detect as many 
buildings as possible. 2™ order errors were accepted as a 
problem to be solved in the following steps of classification and 
building model generation. Hence, the percentage of 2" order 
errors is relatively large. The number of segments, which are 
wrongly classified as house-containing (second order errors), is 
much higher for the rural area than for the urban region. Table 
4-2 contains the statistical figures. The reasons are the general 
characteristics of the locality such as vegetation and height 
differences. Since the detection mainly bases on the degree of 
the variation in height (laplacian values), urban areas are 
segmented more clearly. That means, in dense city areas, that 
the height differences between buildings and their surrounding 
is bigger as in rural areas. Especially commercial units: they 
usually do not have high vegetation areas at all — thus, the 
segments limit the buildings precisely. 
  
Classified correctly 2°“ order errors 
Rural area 770 (94%) 1960 (340%) 
Urban area 1553 (95%) 1114 (170%) 
Table 4-2. Statistic of purely on height information based 
house segment classification 
The exceptional high number of second order errors of the rural 
region can be easily reduced by an analysis of the local standard 
deviation of the height information and the gradient values. 
About 40% of the 1960 second order errors could be 
eliminated. 
Finally, the centre coordinates of the segments selected as 
houses in the house-segment-file are compared with those 
already recorded in the look-up table. If the coordinates differ 
by more than four metres from each other, the new ones are 
added. Segments wrongly classified as house-containing are 
nonetheless included in the house positioning and modelling 
process. Whether there is a house or not, is checked by the point 
distribution. The position and the new outline of the recognised 
building shall then be saved in a new vector file. It is also 
planned to create a model of each house. 
In both building detection procedures, it is important, that no 
segment, which might contain a building, is eliminated. Each 
centre coordinate in the look-up table is assigned a percentage 
expressing the likelihood of containing a house. 
With this method of detecting buildings, very good detecting 
quotes were achieved. In the urban area 99% and in the rural 
area 99.5% of the in 2001 existing buildings were detected. 
5 RESULTS AND CONCLUSIONS 
Based on the procedures described in the former chapters, 99% 
of the buildings in the study area could be detected in the laser 
scanner data. The pixel map proved to provide a good trigger 
for the detection of buildings in the laser scanner data and it 
also decreased the processing time. In addition, a number of 
buildings, which were not present on the map, could be detected 
purely on the basis of the laser scanner data. Due to planimetric 
discrepancies of up to 12 metres, which were detected between 
the map and the laser scanner data set, the final segmentation 
was performed purely on the basis of the laser scanner data, 
without the contribution of the map. The ortho photo was solely 
used for visualization and verification purposes. The problem of 
identifying segments, which do not contain houses, has to be 
addressed in future work on building model generation. 
The use of eCognition proved to be rather beneficial in the 
detection and segmentation procedure as well as in the 
characterization of the segments used for classification. It has to 
be mentioned, however, that the current version of eCognition 
is rather slow in some operations, and that larger regions have 
to be subdivided into small tiles for successful operation. 
6 FURTHER RESEARCH 
Future work should first concentrate on the optimisation of the 
segmentation and detection procedures in eCognition and 
ArcView. Furthermore, methods for processing very large data 
sets have to be developed. Tests with more data sets from 
different sensors and in different regions have to be performed 
to prove the applicability of the developed processing schemes. 
Raw unfiltered first and last pulse data, preferably in 
combination with laser scanner intensity values, should be used 
as input into the detection and segmentation procedure. 
Building modelling will be a major point of future research. A 
discussion of suitable data formats for storing 3-D building 
model information in a topographic landscape model and 
techniques for transferring these models into maps and other 
visualization tools will follow. 
A - 173