6 RESULTS
The image data is taken from a database of aerial im-
ages provided via FTP by the ISPRS Working Group
Il/3, Test on Image Understanding [Fritsch & Ses-
ter, 1994]. Aerial photographs (f—153.19 mm, scale
1:4000) of an urban area were used as test data. The
scanned b/w images (stereo-pair) have a resolution of
60 um per pixel. The image size is 1000 x 1000 pixels.
The images are in epipolar geometry. Camera parame-
ters and projection formulas are given.
In the preprocessing stage for both images about 40000
short lines and 6000 primitive objects LINE are gener-
ated. Using the productions of the production net illus-
trated in Fig. 3 the scene analysis was carried out. As
all generated objects remain stored in the blackboard,
the process of analysis can be examined easily. For in-
teractive selection and visualisation of object sets (in-
terim results) an explanation component is used.
In order to demonstrate the recognition of a single
house all objects generated within the scene section
shown in Fig. 1a,b are illustrated in Fig. 5. Starting
with the objects LINE in both the left and the right
channel the stepwise composition of compatible object
combinations can be traced up to the objects ROOF.
On each stage the image structures are subjected to
additional geometrical constraints by applying produc-
tions. The chaining of productions in a production net
results in logical AND-operations of constraints. Trac-
ing the subimages (L) to (P) in Fig. 5 we realize that
parallelogram-shaped image structures are filtered out
of the primitiv objects LINE.
Fig. 5 shows that both in the left and in the right
channel the number of objects PARALLELOGRAM is
smaller than the number of the objects LINE. Due
to the combination of the objects PARALLELOGRAM
and U_STRUCTURE of the right and the left channel
(P4 in Tab. 2 and Fig. 3) the number of the objects
(RooF.AREA) is much bigger than the number of ob-
jects (PARALLELOGRAM). However, the number of
objects (ROOF) is smaller than the number of objects
(RooF-AREA).
Thus, the structures displayed in Fig. 5 (R) meet the
geometrical relations of the stated productions within
the limits of tolerance set by the parameters. The dis-
played objects Roor cumulate in a small region of the
-scene and are a clear indication of a house. The object
RooF with the best assessment stands representatively
for the house's roof of the scene.
This object is displayed in Fig. 6 (top). The previous
objects which have built up the target object can be
obtained by the derivation graph. A part of this deriva-
tion graph is displayed graphically in Fig. 6. The result
of the analysis of the ISPRS dataset FLAT is shown in
Fig. 7. For a 3D-visualisation of the recognized build-
ings we have assumed a flat terrain and to this plane
the left image was mapped.
836
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Fig. 6: Part of a derivation graph
Bo OD
| Fm
