2e
ct
1,
I-
d
t-
1e
of
to
re
1€
S,
le
1€
is
le
le
ts
ne
le
ce
1d
Its
were measured. Effects on geometric image quality were
up to 2 pixels and, still worse, image distorsions appea-
red to be randomly distributed and unpredictable. Thus
a system calibration in advance of the practical tests
was impossible. Instead, the camera's distortion para-
meters had to be measured during the actual image
acquisition. This was done using the calibration frame
mentioned above, shooting images of the frame imme-
diately before and after the actual images to be evalua-
ted. The two calibration results yielded by bundle-block :
adjustment subsequently were interpolated. This me-
thod may be applicable for all kinds of non-photo-
grammetric cameras, especially unstable digital cameras
with rather fast changing distortion parameters, re-
ducing the work for measuring reference points in the
field. Of course, the calibration object itself must be
reliable and stable enough for this task. If rough starting
values for the orientation parameters can be provided,
targets on the frame can be measured automatically in
digital images since their object coordinates are known.
This method worked with DigiSAS images and can ease
tedious orientation measurements with their many ima-
ge coordinates to be registered.
IMAGE EVALUATION
Orientation and Image Matching
These are standard tasks of digital photogrammetry.
At the TU Berlin the Digital Stereophotogrammetric
System (DSS) generally is used to do the interior, relative,
and absolute orientation as well as to resample the
images to the normal case of sterophotogrammetry
and, finally, to match them automatically in a multi-
step process (Konig ET AL. 1988). However, since almost
all 'real work' in this project was done with images of
a properly calibrated SMK, the actual orientation process
was still easier and consisted of an affine transformation
of each image with respect to the fiducial marks and a
simple rotation/translation to consider offset and tilt
caused by the scanning process. After that, the data
are ready for image matching, i.e. determination of
homologous image points.
A combination of normalized cross-correlation and
least-squares matching is used. Since the latter needs
good starting values, an image pyramid strategy is used
starting at the level of lowest resolution with a one-
dimensional cross correlation process that works with-
out operator assistance. Correlation results are consid-
ered approximate values in the next pyramid level
which makes matching a totally automatic process.
On each level the point raster is densified until the
bottom level is reached. In order to keep computation
time short, only in this original image the least-squares
method is used to provide sub-pixel results. The found
matches undergo a quality control, where several criteria
are tested and it is decided which results are accepted.
Recently, the two steps that consume the lion's share
of computation time, i.e. resampling and matching,
are sped up considerably by the mutation of the DSS
into an Advanced DSS (ALBERTZ ET AL. 1991, ALBERTZ &
Kônic 1991). In the actual constellation this system
uses a 14-transputer Paracom MultiCluster hosted on a
SUN workstation.
DEM COMPUTATION
After the image matching process the parallax file defi-
ning homologous points in the stereo images has to
be transformed into object coordinates. The resulting
3D coordinates are related to the selected geodetic refe-
rence system. The rectangular metal frames around
the soil test areas were used to define a system with a
reference point in the lower left corner (as seen in the
images). This coordinate system allowed the compari-
son of multitemporal images and resulting DEMs, which
was the main objective of the measurements. Before
interpolating a regularly gridded DEM, application of
a filter for the elimination of blunders in the height
points proved to be useful. This filter contained two
components. First, minimal/maximal threshholds for
Z-coordinates (heights) and second, a statistical elimi-
nation of local peaks with extraordinary height diffe-
rences in comparison to their surrounding points. It
was especially designed to eliminate blunders which
could have a negative effect during the following inter-
polation process.
Advantageous for graphical representations as well as
for further numerical analysis is a regular grid of heights.
Several algorithms have been tested with examples of
different relief models, i.e. sharp edges, holes and crests.
The influence of the height point distribution on the
interpolation results was examined because a dense
regular distribution that guarantees the best results pos-
sible often can not be achieved. A good interpolation
algorithm has to be robust concerning gaps in the
distribution which may originate from correlation errors
eliminated by the previous filtering or from shadow
zones where no correlation is possible. The following
interpolation methods were compared: nearest neigh-
bourhood, minimum curvature, polynomial interpola-
tion, and finite elements (HIFI-88).
During tests the minimum curvature algorithm (Brices
1974) proved to be the most suitable method under
given circumstances. HIFI-88 did not perform that well.
However, it was designed to use additional information
(e.g. break lines) in the interpolation process that our
evaluation process did not provide. Polynomial inter-
polation worked well as long as the height points were
densely and regularly distributed, otherwise it had pro-
blems. Nearest neighbourhood is suitable only for rough
reliefs with measured points distributed more densely
than the grid points wanted in the end. The resulting
regular DEM could be processed further by various me-
thods in order to identify and eliminate blunders that
may have passed all previous tests. The last step is the
interactive control and - if necessary - correction of
the DEM which can be displayed in isolines, as a 3D
wireframe model and so on. For special comparison
purposes the extraction of height profiles was imple-
mented and differential DEMs can be computed to
capture relief changes.
