ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision“, Graz, 2002
offset, and the reflectance data for the determination of the
planimetric offset.
Both height and reflectance data can be used for the
determination of the strip offsets, but both have there
advantages and disadvantages. These will be analysed in the
following paragraphs. Furthermore, new procedures will be
suggested to improve the accuracy of the matching and to make
matching possible in a larger number of areas in order to
increase the number of offset measurements between the strips.
2. MATCHING HEIGHT DATA
For stereo matching it is well-known that texture is required in
order to obtain a good precision of the estimated disparities.
Similarly, when matching height data, height variations are
required in order to be able to estimate planimetric offsets.
However, there are restrictions to the kind of height variations
that can be used for matching height data.
* As already pointed out by Kilian et al. [1996] and Maas
[2000], areas that are occluded in one of the strips should
not be used for matching. The usage of heights that are
derived from interpolation over an occluded area results
into systematic errors in the determination of the offsets. In
laser altimetry data sets, such occlusions are mostly caused
by buildings.
Due to the characteristics of the laser sensor, one should,
however, also avoid the usage of height jump edges in areas
that are not occluded. When taking an image with a CCD-
camera of a checkerboard, pixels that cover a part of a white
field and a part of a black field will obtain grey values that
are somewhere in between white and black. Such mixed
pixels do not occur in height images acquired by laser
altimetry sensors. If a laser beam at the edge of a building
roof hits both a part of the roof and the ground, the recorded
height will be either the roof height or the ground height,
depending on whether one selects first or last pulse data.
Hence, the characteristics of a roof edge in a height image
are comparable with the characteristics of an edge in a
binary image. The location precision of such edges, and
thus the precision of matching height images using these
edges, depends on the length of the edges and the
orientation of the edges with respect to the grid [Fôrstner
1986]. For laser altimetry data, the orientation of the grid
corresponds to the orientation of the scan lines of the laser
scanner. In the worst case (edges parallel or perpendicular
to the scan lines) biases of up to 0.5 times the distance
between the laser points may occur in both strips. Hence,
matching such edges in data sets with a point distance of
e.g. 2 meters, may result into an error of 2 meters.
Although the maximum bias that may occur varies with the
orientation of the edges and may average out over a large
number of edges, the quality of the offset estimation is hard to
predict. Whenever possible, one would like to avoid the usage
of height jump edges for the estimation of planimetric offsets,
even if these edges do not cause occlusions. This is in sharp
contrast to matching grey value images where strong step edges
give the best matching results.
In order to estimate the planimetric offsets from the height data,
height variations are, however, required. These height
differences then need to be provided by smooth surfaces with
surface normals pointing in three independent directions. At
least two of these surfaces need to be slanted. Parts of sloped
A - 376
terrain or slanted roof faces can provide suitable information.
Unfortunately, the number of locations in strip overlaps with
such surfaces will usually be small, in particular in rural areas
with relatively flat terrain. Under these circumstances it will not
be possible to find sufficient locations where offsets between
strips can be measured in three dimensions. Therefore, tie
points should be completed with other tie features such as
ridges or planes that only supply information on the offset in
two or one dimension respectively. These dimensions, of
course, do not need to be parallel to one of the axes of the
coordinate system. By combining the different tie features in a
strip overlap, sufficient information should be acquired to make
the strip adjustment possible.
In flat terrain usually provisions are made to drain water. Figure
1 shows an example of such a terrain with many ditches in
meadows and along roadsides. Such linear structures can be
used to determine the offsets between strips in height and in the
direction perpendicular to the ditch orientation.
Figure 1: Height image with ditches.
The height data of such structure can, however, not be matched
with a standard image matching tool. Because of the relatively
small width of the ditches with respect to the distance between
the laser points, not every ditch part is represented well by the
laser points. When computing a DEM from the triangulated
laser points, interpolation between points on either side of the
ditch produces incorrect height values (figure 2). Such errors
would effect the performance of a standard image matching
algorithm.
Figure 2: Perspective view of a DEM part with a ditch.
Viewing the point cloud in the direction of the ditch confirms
that the ditch of figure 2 is uninterrupted. An estimation of the
offset of such a ditch between two strips can be made if the