. Istanbul 2004 
wo orthogonal 
points onto the 
ction the same 
the z-x-plane. 
iddle roof with 
he surfaces are 
e other two in 
ons the three- 
lanes are now 
d in the same 
1e endpoints of 
ne. 
ost-processing 
irfaces can be 
isons are e.g. 
the data set. 
whether they 
recognised as 
the following 
intersecting at 
e on one line. 
. coincide and 
t in a vertical 
from different 
sected as well. 
oof is finished 
  
2 determined, 
eby, a certain 
are engaged 
wall polygon 
it have been 
x-y-plane the 
points of the 
ng model and 
  
   
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
4 RESULTS 
4.1 Assessment 
The quality of the modelling results of the described technique 
depends on the data quality and the complexity of the building, 
which is to be modelled. In the following, the results are 
classified as correctly modelled buildings, partially correctly 
modelled buildings and incorrectly modelled buildings. The 
main reasons for incorrectly modelled buildings are: 
a) Gapsin the point clouds 
b) Strong dispersion of the points due to certain roof 
characteristics or by height misalignment in case of 
multiple flight strips 
c) Buildings that are built into the slope and false 
surfaces that result from this situation 
d) Very small buildings with only few points 
e) Small pitch roofs 
  
  
Figure 3-6: As VRML models visualised buildings 
For the determination of the correctness of the buildings 
generated with the proposed method the Swiss data set was 
used. The probe contains a total of 229 point clouds, 29% of 
the point clouds were not modelled, 9% of the buildings 
were modelled with small errors, and 62% of the buildings 
were modelled successfully. 
The procedure is characterised by a short computation time. 
The computation time of one point cloud with about 300 
points is in average 0.1 seconds using a Pentium 4 (1.6 GHz) 
and 256 MB RAM. 
4.2 Analysis of the modelled details 
The level of detail in the modelled buildings depends on the 
relationship of the feature size to the point density of the laser 
scanner data as well as of the set parameters of the thinning 
procedure. The practical testes have shown that a minimum of 
ten points per plane is required, in order to be able to find a line 
representing the plane. In addition to the detectability of 
  
planes, the definition of the plane outline becomes rather vague 
if only few points represent a plane. 
With a point density of approximate one point per square meter 
(Swiss data set) this means that only surfaces with a minimal 
size of approximately ten square meters can be modelled. A 
larger point density (data set of Freiberg) does not necessarily 
mean an increase of the detail recognizability, since the data in 
this case are also thinned out more in the current 
implementation of the method. 
The procedure tends to a certain generalization. Smaller details 
such as dormers or chimneys are usually not modelled. Still, 
the method is less susceptible to strong dispersions in the laser 
data or insufficient strip adjustment between individual flight 
strips. 
4.3 Comparison with terrestrial measurements 
To determine the geometric accuracy of the modelled buildings 
the coordinates of all comer points of the modelled building 
were compared to terrestrial measurements for these points. 
The mean difference between the modelled and measured 
corner points is + 0.46m in position and + 0.25 m in the height. 
The accuracy in height is better than the position accuracy. 
This is due to the better height accuracy of the laser points. For 
the position accuracy of the modelled corner coordinates the 
major restriction is posed by the point density (average point 
distance approximate 1 m). 
5 CONCLUSION AND OUTLOOK 
The method is suitable for the modelling of the most important 
basic building types as well as for simple combinations of 
those. Advantages of the procedure are to be seen in the 
effective computation and small sensitivity to sub-optimalities 
in the laser scanner data. A wide range of point densities can be 
processed. A disadvantage of the method lies in the necessity 
of thinning out data for proper line detection and the loss in 
small detail associated with it. 
  
Figure 5-1: Example of reconstructed building models in a 
virtual village