ng
ut
ni-
Evaluation and clean-up
Procedures are required that detect the above operating
characteristics.
In order to detect unconnected line-pieces, the skeleton
is vectorized. For each segment the direction is
computed. The decision whether the line segments will
be connected is done by a length-weighted direction
table. The problem is to decide whether fragmentation
of the road boundary is due to the segmentation
technique or due to a new node.
Another procedure has to be defined that detects
parallel boundaries in the region of interest. One
method is to detect jumps in the maximum cost path.
Their detection results in adaptation of the width of
the region of interest and an assumption for the
location of another (parallel) road.
Testing of hypotheses
In fig. 4 three boundaries run over the complete length
of the region of interest, but there is a gap. The
presence of the road is first checked for the parts at
both sides of the gap. Fig. 6 shows how the region of
interest was splitted in two candidate arcs and one
candidate node. The boundaries in the candidate arcs
run over the complete length of the region of interest,
so there is evidence for the presence of a road
segment. However, also canals have parallel borders.
So the conclusion is not unique and further evidence
has to be found by combining edge information with
surface information.
When checking edges in the candidate node, edges
will be found perpendicular to the boundaries in the
candidate arcs. This indicates high evidence for the
presence of a fly-over, but this hypothesis is tested in
another stage of the process.
44 High level: control
After analysis of the segmentation result, the next step
in the hypotheses hierarchy has to be determined.
Control by the high level consists of two parts:
1. local and context confirmation of the candidate road
segment;
2. updating of the evidence of the interpretated
objects.
Local and context confirmation
If evidence has been found for the presence of a road
boundary, it needs to be confirmed by evidence of a
road surface. The hypothesis that the road is still
present is accepted on bases of local evidence. It is
known that roads form a connected network. Hence
we can use context information to find evidence for
this hypothesis. Suppose the evidence of the candidate
node increases in a next step, than the evidence for the
two adjacent candidate road segment increases.
Fig.6 The region of interest of fig.3 splitted in thre
a) candidate road segment; | b) candidate node;
e parts
The problem of the high level is: When is enough
evidence gathered to accept a definitive hypothesis?
For example: How many segmentation techniques
should be used before enough evidence is gathered
about the presence of a road? Ideally, unambiguous
measures, able to compare results of all segmentation
techniques, should indicate the quality of the segmen-
tation result. This measure should not only include
geometry, but aspects like context as well.
Updating of the interpretation evidence
The candidate node has to be examined in a next step
of the hypotheses hierarchy. Because we already found
evidence for the presence of a fly-over, we will first
examine if it is a fly-over. For this confirmation we
need context information too. So recognized parts of
the road network can guide further interpretation.
Future recognized objects may also influence present
decisions. If road segments are found that cross the
current road at the location of the gap, the evidence
that there is a node of type fly-over increases.
5. CONCLUSIONS
Image interpretation of aerial images by computer is a
highly complex task. The casestudy illustrates that
even for one simple road segment in the database, a
complex procedure is needed to verify its presence in
the aerial photograph and to check whether its
properties, such as width and curvature, have changed.
The problem is to bridge the gap between the image,
which is a two-dimensional intensity array, and the
object models.
A multitude of segmentation methods should be
employed in an integrated way. In our opinion first
more insight is needed into the performance of
segmentation techniques on aerial images, before the
image interpretation problem can be solved. A bottle-
neck is the lack of measures for the evaluation and
comparison of the performance of segmentation
techniques.
An inherent problem of interpretation is the integration
of knowledge sources. Based on these knowledge
sources, hypotheses should be constructed, tested and
updated. In a multi-stage approach the parts of the
image that have been examined influence the inter-
pretation process. The evidence of previous interpreted
objects should be updated, with progressing
availability of context information.
Consequently, the computational burden of managing
the interpretation process with its many feedback-loops
is huge.
One of our present points of concerns is whether
expert shells are suitable for managing the complexity
of the interpretation process.
c) candidate road segment.
483