2 VERIFICATION
Verification of GIS data means to find the parts of the data which
have not changed. One way is to compare the results of an auto-
matic road finder with the given data. But this is disadvantageous
because the given data is not used to guide the road extraction and
the matching is computationally expensive. A way to check the
data which avoids this is to compare the data "directly" with the
image data. This has the following advantages: The area to be
investigated is known; there is highly reliable information about
the spatial position and, what is more, about the topology of the
roads because most of the roads normally are unchanged. Using
this information, it is possible to close gaps if they are enclosed
by verified sections. By and large, there is a good chance to verify
roads using a simple model. This section presents an approach
for the verification of GIS data using high resolution image data
(pixel size 10—50 cm) and simulated GIS data representing the
axes of the roads.
2.1 Model and Fundamental Idea
The proposed approach is based on a simple model which com-
prises two fundamental assumptions about the appearance of roads
in aerial imagery: (1) Roads have mostly straight and parallel
roadsides. This means that if a road in the image corresponds
to an axis of the GIS data both roadsides will be approximately
parallel to the axis. (2) Roadsides correspond to strong edges in
the image and the gray values along a road axis are expected to
be more or less constant.
The fundamental idea of the approach is that both roadsides
are close to an axis if the GIS data corresponds to a road in the
image. Therefore, the first step consists in searching for the two
strongest edges at both sides of the axis. This is done with loose
constraints. For that reason some edges which are no roadsides
will be detected. If the axis corresponds to the road the number
of these false detections will be relatively small, otherwise many
randomly distributed edges which are no roadsides will be found.
The decision whether the axis corresponds to the road in the
image is made in the second step using the following criteria:
Straightness, parallelism of the extracted edges, and homogeneity
of the gray values within the expected road.
2.2 Verification procedure
2.2.1 Edge Detection To find the two strongest edges a gradi-
ent image using the modified Deriche edge operator (Lanser and
Eckstein, 1992) is computed. This operator yields good detection
quality, accurate location, few multiple responses, and isotropic
response. Along each axis points with constant distance to each
other are calculated. At these points relatively wide, symmet-
ric profiles are taken from the gradient image perpendicular to the
axis similar to (McKeown Jr. and Denlinger, 1988). The positions
of the two strongest edges within each profile are determined. The
only constraint on the position of the two edge points within the
profile is a minimum distance to each other. In Figure 1a) the
detected edge points are shown as black points superimposed on
the test image (cf. Fig. 3 for the corresponding GIS axes). There
are a lot of outliers due to disturbances near the road.
2.2.2 Width Estimation Because of the outliers in the edge
detection it is important to estimate the actual width of the road.
The center of the two edges and the distance of the center to the
old axis is calculated for each profile. If this distance is less than
a certain threshold (depending on the given level of accuracy),
the two edge points are labeled as roadsides. The longest sections
where the edge points are labeled as roadsides are computed using
the imperfect sequence detector (ISD) described by (Aviad and
54
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Carnine Jr, 1988). For these sections the mean road width is
estimated. After adapting the width of the profiles to the road
width, the search for the two strongest edges is repeated for each
profile. By this means, disturbing edges further away from the
road are eliminated. In Figure 1b) the result after the estimation
of the road width is shown. The benefits of this step can be seen
especially at the curved road in the upper part of the image. In
Figure 1a) (before the estimation of the road width) the edge points
are widely spread, while in Figure 1b) (after the estimation of the
road width) most of the edge points correspond to roadsides.
2.2.3 Evaluation of the GIS Axes Two kinds of errors can
occur when labeling edge pairs: An error of the first kind is
committed if an edge pair is labeled as not corresponding to the
roadsides, although both edge points correspond to them. An
error of the second kind is committed if the edge pair is labeled
as corresponding to the roadsides although this is not the case.
These errors cannot be detected for each edge pair individually.
Therefore, the continuity of extracted edge points is checked along
the direction of the axis.
A frequent reason for an error of the first kind is a slightly
inaccurate position of the axis. This leads to a constant bias of the
GIS axis and the center point of both edges. Therefore, the edge
pair will be labeled as not corresponding to the roadsides. This
error typically occurs for many successive edge pairs. To detect
this kind of error, the string of centers is checked for straightness
along the GIS axis. Each point and its two neighboring points
are connected by two vectors. The criteria for "straightness" are
that the angle between the two vectors, as well as the difference
between the mean direction of the two vectors and the direction
of the GIS axis are small. First, all center points are labeled
individually. Then itis checked if a gap in the string of edge points
preliminary labeled as roadsides can be closed by a continuous
string of center points labeled as straight. If this is the case, the
corresponding edge points are labeled as roadsides as well.
The errors of the second kind are detected by checking all edge
pairs which are labeled as roadsides. This is based on measures for
straightness, parallelism, and homogeneity. Typically roadsides
are straight. Therefore, all edge points which are colinear with
their neighbors are assumed to be faultless, all others to be faulty.
A measure is computed for each roadside separately. To check the
edges for parallelism the direction of the edge points is taken from
the direction image calculated with the Deriche edge operator as
well. A measure for parallelism of the two edge points within
each profile is derived by comparing their directions. It is not
advisable to assume homogeneity of the gray values for the whole
road as there are too many disturbances, like cars or shadows.
However, a great part of the road is homogeneous. What is
more, an area depicting no road will often be distinguished by
inhomogeneous gray values. The gray values of the center points
are accumulated into a coarse histogram. A homogeneity measure
is derived by an investigation of this histogram. The highest
relative frequency will mostly be higher for roads than for other
areas. Furthermore, the number of histogram sections with more
than a certain frequency will be less for roads.
Finally all derived measures are combined to decide whether
an GIS axis can be verified or not.
2.2.4 Handling of Inaccurate Axes At some places GIS axes
don’t coincide accurately with the road axes in the image. Some
parts of a GIS axis lie within the road, whereas other parts do
not. Typically, there is a skip in the position of the edge points
at the intersection of the GIS axis with the roadside. The edge
which is intersected by the axis will be detected continuously,
whereas the corresponding roadside will only be detected if the
axis lies between the two roadsides (cf. Fig. 2). A good hint for
this situation
To detect this
for significan
axis into seve
between the 1
the algorithm
Figure
2.3.5 Dete
the verificati
A new road :
there will be
of new junc
vicinity of tl
two differen
road to find
where it was
evaluates gr
are approxin
the standard
of it is lying
indicates a ji
width. The :
not detectab
possible to «
to the GIS a
has a roadlil
of this brigl
perpendicul
