in the city. 
3 were gen- 
) define the 
he usage of 
e block, in- 
racted from 
orm the 3D 
height val- 
) m for the 
ed on quite 
iding addi- 
pose aerial 
ses and to 
lect model, 
oach linear 
d matched 
iat missing 
4.1 Feature Extraction 
Most existing procedures for building reconstruction make 
use of the assumption that buildings consist of a number 
of straight three-dimensional lines and therefore apply lin- 
ear edge segments as initial primitives to a matching pro- 
cess. We use straight grey-value edges, which are extracted 
by an algorithm proposed by Burns, Hansen & Riseman 
(1986) from aerial images of known exterior orientation as 
primitives of the image description. 
  
Figure 4: Image section with projected ground plans. 
For the verification of the generated 3D building hypothe- 
ses and the determination of the unknown parameter val- 
ues, the extracted grey-value lines have to be matched 
against the corresponding lines of the building model. In 
order to define possible correspondences between image 
and model lines, the knowledge on the position of buildings 
in a global terrain coordinate system, which is provided 
by the GIS in addition to the 2D shape of the buildings 
can be utilized. To determine these correspondences the 
object model and the extracted image primitives have to 
be transformed into a common coordinate system by per- 
spective transformation. For that purpose e.g. grey-value 
lines can be extracted from a stereo image pair of known 
exterior orientation to determine the lines in 3D by stereo 
matching. Alternatively, terrain heights are required in ad- 
dition to the 2D coordinates of the ground plan. In order 
to acquire these terrain heights, existing Digital Terrain 
Models can be used, which are available for many devel- 
oped countries. A DTM can also be computed by standard 
procedures of digital image matching. Problems in build- 
up areas occurring quite frequently due to occlusions and 
height discontinuities while applying these image match- 
ing techniques can be avoided if the supplied GIS data and 
a approximate D'TM is used to mask out building regions 
during the matching process. For our purpose it was suf- 
ficient to use a DTM provided by the State Survey Office 
of Baden- Württemberg (Fed. Rep. of Germany). The 
accuracy of this DTM is in the order of 0.5 m for flat and 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
10 m for mountainous regions. 
Figure 4 shows the ground plans extracted from the ALK, 
which were generalized and projected to the aerial image 
using this authoritative D'TM as data source to define the 
required terrains heights. 
4.2 Matching 
An example for a constructed building, which could be ver- 
ified by matching a 3D hypothesis against the extracted 
image description is given in figure 5. The white wire 
frame shows the initial guess for the building. Therefore 
the ground plan of the buildings were extracted from the 
GIS data base and projected to the image using terrain 
heights provided DTM of the survey authority of Baden- 
Württemberg. 'To construct the building the eave lines 
were assumed to be 6 m, the ridge lines to be 9 m above 
the terrain surface. 
  
Figure 5: Constructed 3D building hypotheses projected 
to image (white) and matched image lines (black) 
  
Figure 6: 3D view of building hypotheses and matched 
image lines 
Straight grey-value lines are extracted from the image by 
the algorithm of (Burns et al. 1986) and matched against 
289