ir endpoints
poundary for
y, we add ar-
he heights of
ilable digital
Fig. 11A,B.
retained for
rimpose the
N the quality
teness of the
B) the origi-
D and (E-F)
rlaid on our
e residential
cted, ten of
teness. The
he algorithm
however, the
ted. The al-
s. The lower
be included
complicated
he right roof
Is to find the
| is correctly
Figure 12: The result of of the 3-D reconstruction on all
houses in the scene of Fig. 5A. The artificial vertical walls are
added and projected down to the ground. The ground height
is estimated through the digital terrain model (DTM). The
marked house is not complete, since two triangular patches
are missing.
7 RULE-BASED SPATIAL REASONING
In a parallel approach [Willuhn and Ade 1996] we want to in-
corporate domain-specific knowledge about houses and house
roofs into the reconstruction process. We think this step is
necessary because, (1) the system should be able to deter-
mine the degree of confidence that the reconstructed object
is really a house and (2) some peculiarities due to practical
or architectural considerations are common in the construc-
tion of houses and should be taken into account. Additional
constraints, such that decisions take place at all levels of pro-
cessing, and that previously executed processes may be re-run
whenever problems at higher levels occur, imply that we need
a system more general than the standard bottom-up. We
propose a system that is capable of iteratively activating pro-
cedures at different levels and based on a uniform knowledge
representation. We have chosen a blackboard architecture
with a semantic network as knowledge representation. Due
to the variety of possible roof shapes, all knowledge has been
coded into rules which have been categorized into the feature,
the structure, and the conceptual level. So far only rules at
the structure level have been implemented. The generated
data from sections 5 and 6.1, i.e. contours, including their
attributes and relations, as well as the 3-D contours and the
planes are used as initial knowledge in the blackboard.
8 FUTURE WORK
Future work of AMOBE includes not only improvement of
each individual component, whenever possible, but also sys-
tem related and conceptual improvements.
For example, we would like to integrate the operator more
actively into the system, especially, for those tasks where the
user instantly can provide approximations, or model or con-
textual knowledge. So far the operator has only been incor-
porated in the building detection phase. This minimal user
interaction works well for the Avenches residential data set,
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
however, in urban scenery fully automatic techniques need to
be augmented with operator guidance, at least in the critical
phases of the processing.
Up to now the color classification and combination with other
cues was performed on only one image. In future investiga-
tions all available overlapping images will be used to test the
improvement of the classification. Furthermore, our investi-
gations indicate that the building detection is better when
more object classes are detected simultaneously. The combi-
nation of multiple cues makes such a detection feasible, and
a possible extension of our research could be in the detec-
tion of all major classes: water, dense forest, separated trees,
grass, bare soil, roads and other paved spaces, buildings and
shadows. The detection of just trees, buildings and water is
important for the reduction of a DSM to a DTM.
The interaction between 2-D and 3-D processing has proven
extremely useful, however, its full potential has not yet been
investigated. Closely related to the interaction between 2-D
and 3-D is the explicit or implicit use of object models. The
issues of object modeling has to be investigated further [Ma-
son 1996]. In future work we will validate our algorithms on
other data, such as industrial and dense urban scenes. We
also plan to improve the data flow by integrating the individ-
ual software modules under one joint system.
9 CONCLUSIONS
In this paper we have presented our strategies, the current
status of research, and made an outlook onto future work.
In the project, we have focused on the 3-D reconstruction of
residential houses, as being the most prominent man-made
objects in high-resolution aerial images. The approach is
highly data-driven, exploits both 2-D and 3-D processing, and
reconstructs roofs of houses directly in 3-D. This approach
has proven powerful enough so that, in contrast to other ap-
proaches of generic roof reconstruction, we can handle more
difficult and varying houses.
We have further shown how digital surface models and color
classification can be combined to detect buildings and in ad-
dition, to provide a coarse description of the buildings. As an
alternative approach to house reconstruction, we have also
reported on a rule-based system, which is built on a black-
board architecture.
The current status of AMOBE is indeed promising and future
undertakings will most certainly profit from the ideas and
results presented here.
ACKNOWLEDGEMENTS
We acknowledge the support given to this research by ETH
Zurich under project 13-1993-4. The authors would like to
acknowledge the work of A. Sibiryakov on color segmentation
and blob analysis. We further appreciate the support of P.
Fua at SRI International and M. Stricker at IKT, Zurich.
REFERENCES
[Baltsavias et al. 1995] E. Baltsavias, S. Mason, and
D. Stallman. Use of DTMs/DSMs and Orthoimages to
Support Building Extraction. In A. Grün, O. Kübler, and
P. Agouris, editors, Automatic Extraction of Man-Made
Objects from Aerial and Space Images, Birkhauser Verlag,
Basel, pages 199-210, 1995.
329
