he theory of
on to further
bility assign-
is way results
oilistic meth-
at in case of
bility models
ve Dempster-
ally updated
oss( B))
oss( B))
to probabilis-
ependent, i.e.
B are mutu-
hould be zero
ROADS
ith automatic
oad database
/ uncommon.
extraction to
publications,
incertainty in
any areas of
r to seriously
| between the
iting process.
shown and it
eps was dealt
e contents of
| first two are
o papers was
same strategy
iss the advan-
gy for updat-
d in the new
tion between
artificial road
ation is lower
roblems arise
side the road.
wledge about
esult. Hence,
erification re-
of the road's
inear features
are performed
h some effort
S, conditional
Figure 3: Verification and detection of junctions [Gunst,
1996]. In black are detected changes. Starting from the
middle of the changed area, new roads are detected and fol-
lowed.
probabilities may be derived that can give these tests a sta-
tistical basis. The question, however, remains whether a test
on linear features can be sufficient to confirm that the road
has not been changed. In other words, do extracted edges
that match in position with the roadsides in the GIS indeed
represent the sides of a road, or can they be due to some
other linear structure? This question is important for auto-
matic updating, since a false verification will not result in a
search for changes in the road network and therefore cause
omission errors in the updated database.
Nevatia and Price [1982] use a relaxation labeling technique
to match a structural description of a map to a structural
description extracted from an aerial image. The nodes in
such a graph-like description do not represent propositions
that are either true or false, but the different assignments of
map features that can be made to a certain feature of the
image. In [Price, 1985] different relaxation schemes for up-
dating the likelihoods of assignments are compared. Although
some schemes are called probabilistic, the final likelihoods of
the assignments cannot be considered as probabilities. Since
much of the evidence found in the labelings at the neigh-
bouring nodes is used several times in the updating process,
many labeling "probabilities" converge to either zero or one.
The mapping between the descriptions that follow from the
likely assignments is, however, very useful for hypothesizing
the correct correspondences.
Haala and Vosselman [1992] use relational matching to match
the structural descriptions of a map and an image. Their
evaluation measure is based on the mutual information be-
tween the two descriptions and is derived from conditional
probabilities that were obtained in many feature extraction
experiments. Using the same probabilities the distribution of
the mutual information was also derived. Thus, it is pos-
sible to define a statistical test on the amount of mutual
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
information of the best mapping found. Since the structural
descriptions contain many attributes of features and their re-
lations, the probability of accepting a wrong match is fairly
low. Although the acceptance test was well defined, the num-
ber of performed matching experiments was too low to draw
conclusions about its applicability.
Summarizing, one may conclude that in many verification
tests it is possible to base these tests on a statistical analysis.
However, verification errors are likely to happen in case of
unmodeled occluding objects or poor object descriptions.
6.2 Detection and measurement
Gunst and Hartog [1994] and Gunst [1996] consider the case
of detecting and mapping new exit roads and fly-overs. After
the verification steps several locations with a possible change
are marked. In these areas goal-directed segmentation al-
gorithms try to detect parts of other roads of the junction.
Once these have been found, the new road parts are classi-
fied as either an exit road or a fly-over. The decision is based
on the values of a few attributes (e.g. the angle between
road elements) and knowledge of the road design rules (e.g.
exits only have a small angle with the main road). Probabil-
ity distributions of the attribute values are not used in this
test. They could, however, have been used to show the un-
certainty in the classification. In case of a high uncertainty it
would then be useful to consider multiple hypotheses about
the kind of junction. In the current implementation no alter-
native classifications are considered.
Cleynenbreugel et al. [1990] suggest to use many more layers
of a GIS for the detection of new roads. Except for old roads,
other information like land cover, DEM'’s, and hydrological
information can also be helpful. Since land cover will dis-
criminate between urban and rural terrain, expectations for
the shape of road networks can be tuned to these classes.
DEM's can be used to derive slope maps that constrain the
possible directions of roads. It may even be possible to derive
probability distributions of the road direction at some point
given the terrain slope at that point. Finally, hydrological in-
formation (position of rivers, lakes, etc.) is also useful. Roads
in mountainous areas are often parallel to rivers and have as
few bridges (=construction costs) as possible. Roads in the
middle of lakes are very unlikely. Many of these heuristics are
valuable for image interpretation and will indeed be used by
human operators.
McKeown and Denlinger [1988] use profile matching for
tracking roads. Road trackers are usually initialized at po-
sitions indicated by an operator. In case of updating road
networks, it can, however, also be done automatically at those
positions where new junctions have been found. Vosselman
and de Knecht [1994] use the least squares method for pro-
file matching and Kalman filtering to estimate the position,
direction and curvature of the road. This approach enables
them to also estimate the precision of the road parameters
and to detect failures in the profile matching. Thus the un-
certainty in the road extraction is fairly well described. Road
trackers, in general, can however only deal with simple roads
and will fail at e.g. Y-junctions.
Other methods to outline roads are often based on snakes,
deformable templates or dynamic programming. Grin and
Agouris [1994] combine the advantages of snakes and least
squares template matching by constraining the matching re-
sults. Precision estimates are also obtained.
915
los al
