m 
age 
Iso 
ch- 
ft- 
to 
art- 
ind 
on, 
tis 
ion 
um 
tos 
are 
ted 
ion 
ion 
of 
the 
es. 
ich 
er- 
the 
ses 
ot 
re- 
e When applying MATCH-T to large scale and high res- 
olution imagery, the percentage of points matched on 
roof tops or trees is very high, resulting in a surface 
description which includes these objects, especially in 
very densely built-up areas. 
e The regularization, even if using local adaptive tech- 
niques (cf. WEIDNER 94), has a tendency to smooth 
the surface and to round off sharp edges. This is a 
big disadvantage, especially when using this kind of 
surface description for building detection. 
e The strategy of MATCH-T is based on matching only 
two images. Our approach is to use four images show- 
ing the same ground area. To our knowledge there is 
no commercial DTM generation package available that 
uses simultaneously more than two images, except for 
combining the final DTM's, which is not an appropri- 
ate approach to use the information. 
  
Figure 4: Basic principle of automatic DTM generation 
Figure 4 shows the principle of deriving a DTM from a set 
of 3-D points computed by feature matching and the effect 
of the regularization. The solid line represents the real sur- 
face, the dots show the 3-D point cloud including some false 
matches, and the dashed line the surface description derived 
using a regularization technique. 
4.2 Using morphologic filtering to determine the topo- 
graphic surface 
As shown in the last section, applying regularization tech- 
niques directly to the point cloud is not an appropriate ap- 
proach to derive a description of the topographic ground sur- 
face. 
Therefore, we first use a special filtering technique to elim- 
inate points matched on buildings, and use the filtered 3-D 
point cloud to estimate the regular DTM grid. This filtering 
is applied to the combination of all six 3-D point clouds de- 
rived from imagery showing the same ground area. The basic 
idea behind the filtering is the fact that buildings or trees 
are higher than the surrounding topographic surface within a 
certain region. 
A similar filtering technique is used by FORSTNER, WEIDNER 
95 for purposes of building extraction from high resolution 
DTM's. They apply a morphological filter on a regular DTM 
grid to derive an approximation of the topographic surface 
without buildings. The difference in our approach is that we 
apply the filtering directly to an irregular set of 3-D points, 
and we added some robust techniques to prevent from falling 
into local minima due to outliers i.e. false matches. 
The principle of the filter we use is from the field of mathe- 
matical grey scale morphology (cf. HARALICK ET AL. 87) 
called opening. 
This generally is an erosion 
z=zOQuw (1) 
followed by an dilation 
Z=20w (2) 
both using the same structural element w(z,y) having con- 
stant z values. The structural element generally may have an 
arbitrary shape depending on the application. 
In order to apply this theory to the irregularly distributed 3-D 
point cloud z(z, y), to eliminate points on top of buildings or 
trees, we first generate a regular grid h(z, y) with a fixed grid 
size ó. One has to make sure that the structural element W 
is not entirely contained in a building outline. On the other 
hand, it should be small enough to keep small hills in the data 
set. We choose a squared window of the size of 200 f. With 
that, the opening can be performed by simply placing the 
structural element on each grid point and applying a robust 
minimum filter 
bos int { (eles c wh (3) 
followed by a maximum filter 
= sup {Ale leg € wh (4) 
As the 3-D point cloud contains a certain percentage of out- 
liers, it is not appropriate to use the absolute minimum for 
the erosion (minimum filtering). This would cause gross er- 
rors in ki. Therefore, we use a generalization of the median, 
the so called k-th sorted element (k = 50% is equivalent to 
the median). The robust property of the generalized median 
is very useful to eliminate outliers i.e. points that are below 
the ground surface. From numerical investigations, we found 
k = 3% appropriate for the erosion. 
The opening results in a regular grid h(z,y), which is a dis- 
crete approximation of the topographic surface. Using this 
surface description, each point in the 3-D point cloud can be 
classified as being: 
above the topographic surface, 
(Mainly points on top of buildings and trees, but 
also some outliers) 
on the topographic surface or 
below the topographic surface 
(Mainly outliers due to false matches) 
by applying the following thresholding scheme 
above : (z(ny)-h) » Ah 
As,y)=4 e : |ze9-R| s A^ (9 
below : (z(z,y) — kh) « —Ah 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996