oof detail. 
ri 
bof detail 
-ABal]. 
in the fol- 
e digitized 
| to chose 
olution of 
rammetric 
the world 
tered with 
rithm and 
al. 1986] 
  
ta for 
ation sys- 
nt of each 
s of build- 
f outlines' 
Aonocular 
s footprint 
| matching 
ne match- 
ed succes- 
zment and 
two neighbours. For this case we can generalize perspective 
to affine transformation. The start values for this procedure 
are extracted from the set-up data of the aerial photograph 
and the known coordinates of the map data, combined with 
at least approximate elevation data of the ground and ap- 
proximate knowledge of the height of the building. Results of 
the matching procedure of GIS-building's outline and Burn's 
lines are presented in Figure 5. 
LL y 
Figure 4: Four steps of piecemeal affine matching in 
the case of a rectangle (4 lines outline). 
  
  
  
  
  
Figure 5: Results after piecemeal affine match- 
ing (based on distance transformation) between GIS- 
building's outline and Burn's line segments. 
3.4 The 3D roof skeleton (Stereoscopic approach). 
The results of the fusion process are enhanced and verified 
by the involvement of a second image, which allows a 
stereoscopic investigation of the scene. Starting from 
the extracted outline of the roof we create a fully three- 
dimensional set of roof lines, which we call the roof-skeleton. 
This data set has to be topologically checked and finally 
leds to the CAD-model of the entire roof. A number of 
different steps of the procedure is presented in Figure 6 and 7. 
4 RESULTS FROM TEST SITE GRAZ 
The medieval roofs of downtown Graz give us the background 
of a suitable test site for our investigations. Various details 
and subdivided roof shapes measure the quality of our ap- 
proach. The current result of the algorithm we are working 
on is depicted in Figure 6. 
From the token-set of lines in Figure 3, derived from digital 
image data we start to extract the roof outline (6a) and ex- 
ploiting hypothesies of angularities between roof outline, ridge 
lines and other edges of the roof we create the roof skeleton 
step by step (6b,6c,7a,7b) (cf. parsing in [Stokes, 1992]). 
The final result, which needed a small amount of manual 
interaction is compared with the digital image of the scene 
(7c). 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
Figure 6: Different steps of the extraction of the roof 
skeleton based on image data and 2D GIS (6a, 6b, 6c). 
5 FURTHER INVESTIGATIONS. 
The result of our procedure may be understood as an initial 
step towards automated roof reconstruction, which needs to 
be expanded and improved. A multi-image approach shall be 
one of our further investigations. This may cause a better 
accuracy by means of least squares adjustment, available 
from multi-stereo solutions. Beside the improvement of the 
quality of the derived roof skeleton, we have to increase the 
level of detail of the CAD-model by means of detection of 
chimneys, sky-lights and other small parts of roofs. This is a 
must if phototexture is involved to enhance the CAD-model 
of the roof in order to guarantee the correspodence between 
phototexture and geometry [Gruber et al. 1995b]. 
We also intent to show, how a verification process of 
the automatically derived CAD-model of the roof may be 
lead by image processing methods. This means, that the 
correspodence between CAD-data and texture-data may be