Summarizing, relatively simple tasks like outlining single, and
completely visible roads can be performed successfully. If
the algorithm can be formulated as a least squares estima-
tion problem, quality descriptions can also be obtained. The
more complicated tasks that relate to interpretation of junc-
tions and the road context are far from being solved. It is
also clear, however, that much of the available knowledge is
not adequately modelled and, therefore, not available to the
interpretation process.
7 DISCUSSION
The main issues raised in this paper were the different ways to
describe and process uncertain data related to updating road
databases. It was also shown that there are many reasons for
the poor results in image interpretation.
Many ways to describe and process uncertain data are based
on probabilities and Bayes’ rule. Several alternatives, like cer-
tainty factors, Dempster-Shafer theory, and possibility theory,
have been applied successfully in some domains of Al, but
may lead to wrong results in other applications. The need
for conditional or prior probabilities is often mentioned as a
disadvantage of Bayesian probabilistic reasoning, but, in gen-
eral, the alternative strategies can not fill the gap of such
a lack of (modeled) knowledge. Gathering this information
remains important. Many claimed advantages of alternatives
for probabilities, e.g. possibility theory, can also be realized
with a probabilistic approach [Cheeseman, 1984].
Probabilities describe uncertainty, but probability numbers
themselves are also uncertain. Especially intuitive heuristics
will often only give us a rough idea about some probabil-
ity number. The endorsement theory [Cohen, 1985] studies
how to represent and reason with heuristic knowledge about
uncertainty. An interesting analogy can be found between
the concept of external reliability in least squares adjustments
(the influence of an error in an observation onto the estimated
unknowns) and the question how changes in probability dis-
tributions affect the outcome of a reasoning process.
Most algorithms for automatic road detection have very low
success rates. The context of an old road database contains
very useful knowledge to improve this. However, this knowl-
edge is far from sufficient to solve image interpretation tasks.
Much additional knowledge concerning the appearances of
roads in aerial images and the context of roads will need to
be modelled.
Uncertainty plays an important role in this knowledge. Much
of our knowledge is heuristic and therefore uncertain. This
uncertainty needs to be described in order to properly reason
with knowledge. In many cases conditional probabilities will
be appropriate, e.g. P(road direction|terrain slope).
Propagation of errors is also necessary. The purpose of error
propagation is not only to assess the quality of the final re-
sult, but also to value the correctness of intermediate results.
The latter motive may be even more important. A sound in-
terpretation can only be made when the quality of all data in
all processing steps is known. If the results of some step are
found to be uncertain, this knowledge can be used to formu-
late multiple alternative hypotheses instead of only pursuing
the most likely one.
Modelling knowledge and propagating uncertainty are two
complicated tasks. A lot of effort will be required to ob-
tain satisfactory results in automatic image interpretation. It
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
therefore seems a good approach to first start with semi-
automated methods and gradually increase the interpretation
tasks of vision algorithms.
LITERATURE
Baumgartner, A., C. Steger, C. Wiedemann, H. Mayer, W.
Eckstein, and H. Ebner [1996]: Road Extraction for Update
of GIS from Aerial Imagery: A Two-Phase Two-Resolution
Approach. International Archives of Photogrammetry and
Remote Sensing, vol. 31, part B3.
Blahut, R.E. [1987]: Principles and Practice of Information
Theory. Addison-Wesley, Reading, MA, U.S.A..
Buchanan, B.G. and E.H. Shortliffe (Eds.) [1984]: Rule
Based Expert Systems. The MYCIN Experiments of the Stan-
ford Heuristic Programming Project, Addison-Wesley, Read-
ing, MA, U.S.A..
Cheeseman, P. [1986]: Probabilistic vsFuzzy Reasoning. In:
Uncertainty in Artificial Intelligence, vol. 4, L.N. Kanal and
J.F. Lemmer (Eds.), North-Holland, Amsterdam, pp. 85-102.
Cleynenbreugel, J. van, F. Fierens, P. Suetens and A. Ooster-
linck [1990]: Delineating Road Structures on Satellite Images
by a GIS-Guided Technique. Photogrammetric Engineering
and Remote Sensing, vol. 56, pp. 893-898.
Cohen, P.R. [1985]: Heuristic Reasoning about Uncertainty:
An Artificial Intelligence Approach. Research Notes in Artifi-
cial Intelligence, Pitman Publishing, London.
Egenhofer, M.J. and R.D. Franzosa [1991]: Point-set Topo-
logical Spatial Relations. International Journal of Geograph-
ical Information Systems, vol. 5, pp. 161-174.
Fine, T. [1973]: Theories of Probability. Academic Press,
New York.
Fuchs, C., F. Lang, and W. Forstner [1994]: On the Noise
and Scale Behaviour of Relational Descriptions. International
Archives of Photogrammetry and Remote Sensing, vol. 30,
part 3/1, pp. 257-264.
Garnesson, Ph., G. Giraudon, and Montesinos [1990]: An
Image Analysis System, Application for Aerial Image Inter-
pretation. Proceedings 10th IAPR International Conference
on Pattern Recognition, Atlantic City, vol. 1, pp. 210-212.
Goedhart, B. [1994]: A General Framework for Compositional
Reasoning with Uncertainty. PhD-thesis, Delft University of
Technology.
Goodchild, M. and S. Gopal (Eds.) [1989]: The Accuracy of
Spatial Databases. Taylor & Francis, London.
Grin A., and P. Agouris [1994]: Linear Feature Extraction
by Least Squares Template Matching Constrained by Internal
Shape Forces. International Archives of Photogrammetry and
Remote Sensing, vol. 30, part 3/1, pp. 316-323.
Grin A. and H. Li [1994]: Semi-automatic Road Extraction
by Dynamic Programming. International Archives of Pho-
togrammetry and Remote Sensing, vol. 30, part 3/1, pp.
324-332.
Gunst, M.E. de, and J.E. den Hartog [1994]: Knowledge-
Based Updating of Maps by Interpretation of Aerial Images.
Proceedings 12th International Conference on Pattern Recog-
nition.
Gunst, M.E. de [1996]: Knowledge Based Updating of Road
916
Map
Tecl
Gup
Spat
Haal
river
of P
pp.
Heu
agat
Jour
322.
Heu
spat
Hun
Rela
Ana
Kan
Inte
Mai
aeri:
Pro:
Mar
Verl
Merk
ods
Com
pp.
Nev
Aeri
Mac
Nils
gen:
Pea
tem:
Pub
Pric
pari
Inte
Ros
Lab
Syst
Sha
Prir
Sha
The
Vos
in C
Vos:
prof
Wor
fron
264
Wal
drav
