23 
  
  
ES S Bad 
Figure 5: Example of filtered 3-D point cloud superimposed on a stereo model. Black crosses mark points that are removed, 
white crosses are kept. Remember: The point cloud is generated using imagery from six different models, therefore some 
points may look erroneous. 
The threshold parameter Ah can be derived from prior 
knowledge about the height of buildings taking the accuracy 
of the 3-D point cloud into account. We use 5ft. The 3-D 
points classified as being on the topographic surface can 
now be used to estimate a high precision DTM, especially 
in urban areas. This represents the real topographic surface 
much better than a DTM estimated using all 3-D points 
from the point cloud directly. 
Figure 5 illustrates an example of the above described filter- 
ing. The image scale was 1 : 10, 000, the image is scanned at 
60pm. Black crosses are classified as being above or below, 
and white crosses are classified as being on the topographic 
surface. 
One can see that the 3-D points classified as being above are 
mostly on top of buildings or on trees. It is remarkable that 
the outline of buildings and even structures of roof construc- 
tions are recognizable in the classified 3-D point cloud. Fine 
structures of roofs may be recognizable in a higher resolution 
imagery, say 15jym at the same image scale. It seems to be 
very promising to use additional structural and radiometric 
information about buildings to separate points on trees from 
those on buildings and to extract the 3-D structure of the 
buildings using the 3-D point cloud. 
4.3 DTM generation in a production environment 
The above described technique has been embedded in a pro- 
duction environment, so that the DTM generation, works 
fully automatically in batch mode for a whole block of im- 
872 
ages, the number of images being limited only the available 
disk space. The whole procedure of automatically generating 
a DTM for a whole block of images consists of the following 
steps: 
1. Automatic model setup. 
Within this first step the overlapping area of all the im- 
ages is determined by projecting the image boundaries 
onto the ground using a coarse DTM. Each overlap- 
ping area, which fulfills a criterion based on the base 
length of the two images, is declared as a model area 
and the necessary project files for running MATCH-T 
are automatically generated. 
2. Generation of the 3-D point cloud using MATCH- 
T 
For each model found in 1) the 3-D point cloud is 
generated in batch mode. Only the model area is pro- 
cessed which saves disk space and computation time. 
The result of this step is a 3-D point cloud for the 
whole project area which is, except for the project bor- 
der, generated from four different perspectives, respec- 
tively six different model combinations including the 
diagonals (which only occurs in 60/60 geometry). 
3. Point cloud filtering 
From the 3-D point cloud, covering the whole project 
area, so called DTM units are extracted and processed 
by the morphologic filter. The result of this step is a 3- 
D point cloud which ideally includes only those points 
on the ground, all points matched on top of trees or 
buildings are eliminated. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996 
  
5 
p