to the question to what extend information has to be
available to model discontinuities within the object, in
order to garantee acceptable results.
The collection of the informations needed might be
achieved by visual inspection of the image data or by
use of other sources (information systems, maps etc.).
The types of functionals as the size of the regions, for
example, could be determined by some rough knowled-
ge about the topography. The definition of the region
borders will be done by some measurements within the
digital images (polynomials, corner points etc.). The
same is valuable for the discretization of discontinuities.
Much more demanding but of greater importance for
future developments is the attempt to extract these
informations automatically from the image data availa-
ble. The realization of this approach would equip the
algorithm with interpretative capability, what for sure is a
necessity to attain the required flexibility within fully
automated procedures.
One important prerequisite for the success of the pre-
processing is the existance of some dependencies
between the distribution of the image densities and the
geometry of the corresponding surface part. Although
there won't exist direct relations, this assumption will
hold to a certain degree. This might be illustrated by the
effect of illumination with direct solar radiation, which
determines the distribution of enlighted and shadowed
areas by the geometric disposition of the surface. Furt-
her informations like brightness, texture, edges etc. are
depending on the underlying objects too. This data
allows to differentiate between objects, what in many
cases will be equivalent to a discrimination of changes
within the surface morphology.
Existing approaches /Besl, Jain 1988; Grimson,
Pavlidis 1985; Haralick et al 1983/ document that the
interrelations between density values and surface topo-
graphy might be useful to approximate the surface or to
segment the images into homogeneous regions. Furt-
hermore there are trends towards algorithms performing
an extraction of complete objects /Fórstner 1991/ which
likewise could contribute to the evaluation of the requi-
red informations.
Once, the additional knowledge is available it will be
used within the algorithm to control the calculations. This
might be done by selection of those points within a zone
of homogeneous morphology which have to be combi-
ned to a region, or by evaluation of the results with
respect to the type of surface.
In addition the knowledge may be used to segment the
surface elements into different groups. This is useful in
case of discontinuities, because then the surface
elements within a region may belong to different sets of
surface parameters. Similarily the grouping is conve-
nient to exclude elements from the calculations or to
weight them down, what is necessary to eliminate or
reduce the contribution of some surface parts.
FIRST RESULTS
The presented conception has not yet been realized
131
completely, but at least the principle of the object space
based correlation is already transformed into a computer
algorithm. However, this algorithm does not yet have
tuning or control capabilities as they will be necessary to
optimize the results. It therefore seems to be likely to
expect further improvements with increasing complete-
ness of the algorithm.
The tests have been done with already used material
/Boochs 1987/ showing a topographic surface projected
into a medium scale (1:12000) image pair. The digital
images have a resolution of 50 pm. For control purposes
a set of manual height measurements is available.
The tests are performed within three small object areas,
each of them exhibiting different morphology (hillside,
urban, meadow).
The results are compared with calculations from a diffe-
rent correlation algorithm (ARCOS), which has been
proved as valuable for the determination of digital eleva-
tion models of smooth topographic surfaces.
Table 1 shows the mean square residuals (m ) from the
differences between correlation results and manual
measurements, the mean square residuals after blunder
detection (m) and the number of blunders (out).
obj.sp.cor. ARCOS
m m out m m out
qr qr qr qr
meadow 0.27 0.15 1 0.14 0.14 0
hillside 0.23 0.19 1 0.68 0.53 1
urban 0.99 0.99 0 2.41 0.71 5
Table 1: Mean square residuals [m]
Regarding at the area with the simplest morphology
(meadow) we find, that the general accuracy (m) is
equivalent to the known one from ARCOS and seems to
be acceptable (0.07*/.. h ) for this application. Certainly,
a blunder appears within the object space based correla-
tion. However, it will easily be detected and substituted
when the internal program control is available.
Considering the results from the hillside, a considerable
improvement has to be found. This reflects the advanta-
ges of a higher complexity within the functional descrip-
tion of the surface. In this case a polynomial of order 4
has been used.
Finally, even for the urban area some positive aspects
may be found. Although m decreases (0.71m->0.99m)
there arise no more blunders leading to an improvement
in m . Nevertheless, these results are not satisfying.
Further enhancements are to be expected, when the
discontinuities within urban regions will be modelled
correctly. In the present stage with the use of polyno-
mials, results can't be better. After all, the use of small
and fix image windows seems to be valuable, as the
geometric distortions will decrease, what might be an
explanation for the disappearence of the blunders.
CONCLUSION
The presented conception has to be seen as contributi-
on towards higher performance and flexibility of automa-