CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
Figure 9. DSM draped with ortho-image, representing an 
industrial area at the urban fringe. 
5. SPATIAL FILTERING 
To further improve the global quality of the surface model and 
especially to reduce smoothing effects, spatial filtering is 
applied on the height values of the DSM. In a first approach, an 
order statistics filter is applied on the surface model. More 
specific a small 7 by 7 median filter is used, which not only 
reduces noise and outliers but also enhances edges. The value of 
each pixel is changed by looking at the surrounding pixels 
within the 7 by 7 kernel and arranging all values in sequential 
order. Next, the 50 th percentile value is assigned to the centre 
pixel. As the median value is assigned, the influence of outliers 
within the moving window will be reduced. The outcome of 
applying a median filter on an urban surface model is further 
discussed in (Jacobsen, 2006). 
Figure 10. Position and orientation of profile A through 3 
similar buildings. 
Figure 11. Graphs illustrating profile A before and after median 
filtering of the 3 m resolution tri-sterescopic surface model. 
After median filtering, local variations and outliers are reduced 
and the rooftops are at a more or less constant level. 
A method is also developed to further improve building shapes 
based on the knowledge of building contours. Flat roofs can be 
assumed for the buildings within the study field. A first attempt 
failed, where the matched edges were used as approximations 
for building contours. As can be derived from figure 5, the 
extracted edges are not closed polygons and sometimes they are 
connected together with edges of neighbouring buildings. This 
made the conversion to individual building contours extremely 
complex. 
The results of a second approach are more effective. An external 
dataset is used, consisting of 2D building footprints which were 
plotted on aerial imagery by IMP-Bimtas for cadastral purposes. 
Fitting of the 2D building footprints on the generated surface 
models, allows to extract all man-made objects. Subtraction of 
the DSM with the generated building model results in a terrain 
model (DTM) with gaps where the buildings were positioned. 
Distinction between a terrain model layer and a building model 
layer allows to apply different spatial filters adapted to the 
specific needs of the layer. The terrain model without man 
made objects should be a continuous and smooth surface. As 
smoothing constraints are very important for the DTM, a 
median filter with a large kernel size of 18 by 18 pixels is used. 
On the other hand, smoothing must be minimized for the 
building layer to model shape and discontinuities of man-made 
objects as good as possible. As the “bell-formed” shape of 
buildings in an unfiltered surface model is mainly an 
underestimation of height, an upper quartile filter with a small 
kernel of 7 by 7 is applied on the building layer two times 
within the boundaries of each footprint. An upper quartile filter 
is a nonlinear, order statistics filter and returns the 75th 
percentile value within the kernel. Spatial filtering of the height 
values within each building footprint reduces the local 
variations and puts the roof height on a more or less constant 
level. In a final step the DTM is merged with the building layer 
to obtain a final filtered DSM. 
94