What
that
to the
t this
major
ored.
o not
have
of the
Je) is
)osed
been
itions.
. This
ark, to
same
latest
digital
ement
ms to
nputer
s also
e and
close-
1991).
rred to
gorith-
es for
nming
juares
aruen,
can be
an be
oak or
ssible,
either.
gue of
opriate
cation.
fierent
yvalue
values
keeps
ch that
ed with
ry form
cting a
st. The
model
hat the
. More
traction
sed for
model-based measurement both in close-range (Chapman
et al., 1992, Streilein, 1994, Li, Zhou, 1994) and aerial
(Quam, Strat, 1991, Lang, Schickler, 1993) applications.
A first step into the right direction is realized in the
PhotoModeler system (http://www.photomodeler.com.),
where a coarse CAD-model of the object is used to
determine interactively the approximate values for the exte-
rior orientation of the images. This is achieved by
backprojection of the wireframe model of the object into the
images and by operator-guided matching of the model
projection and the respective image features.
Atthe next level, the measured feature primitives have to be
structured into 2-D or 3-D objects for further use in CAD
systems and GIS. This structuring also needs support from
automation, because manual procedures are very time-
consuming. Very little is known about related functions on
Digital Stations.
3.4 DTM generation
The problem of image matching for precise and reliable
DTM generation is not solved yet. This holds for academic
approaches and even more for commercial software.
Occasionally users of automated DTM generation software
remark that the ,results are not very good, but acceptable
for orthoimage production.” This is a comment born out of
frustration rather than a convincing argument. In the past,
DTMs have been created to such a level of quality that they
could be used for many different purposes. Actually, among
all geo-related data sets a DTM was considered the most
permanent and reusable set over time. Are we satisfied
nowadays in producing throwaway DTMs*, just good for
orthoimage production at a particular scale? Can we accept
DTMs which cannot produce ,good looking contours" and
which deliver, if at all, output statistics which „are not helpful
enough“ (Torre, 1996).
In brief the major problems with automated DTM generation
software are:
- Recognition and measurement of object edges and
geomorphologically important features
- Bridging of regions with poor signal content
- Handling of occlusions and shadow areas
- Reduction of a Digital Surface Model (DSM) to a
Digital Terrain Model (DTM); this includes recognition
of trees, bushes, buildings, etc.
- Quality assessment; internal quality control (blunder
detection and location)
We have tested commercial DTM software (Leica/Helava
DPW 770, Virtuo Zo) on various projects under varying
conditions. Detailed reports can be found in Brossard,
1994, Baltsavias et al., 1996. Without manual editing, the
results were not convincing.
In summary it must be noted that to achieve good quality
results which are equivalent to operator measurements, a
substantial amount of editing is necessary. We believe,
however, that there are concrete possibilities to improve the
results. One is to use a multi-image approach, as empha-
sized in the following. The other is to use color and texture
measures for tree, bush and building detection and
separation (Henricsson et al, 1996). Thus the purely
geometrically based reconstruction procedures could be
131
supported by image understanding algorithms, which are
currently investigated at various research labs.
Since DTM generation is basically an ill-posed problem, a
remedy cannot consist in producing a great number of
points for which no reliable quality measures are available.
At least the problem of reliability could be tackled by using
more than two images in the matching procedure. This
multi-image mode has been suggested as early as in
Gruen, 1985a and the first results of this approach have
been presented in Gruen, Baltsavias, 1986. Recent inves-
tigations with a modified version of this concept confirm the
good performance (Maas, 1996).The multi-image mode
improves precision and especially. reliability significantly.
Depending on the number of images used simultaneously
in matching, the number of blunders can be reduced
dramatically. Also, occlusions can be handled very well with
this approach.
3.5 Triangulation
Many optimistic statements have been given recently by the
user community concerning the advantages of digital and
(semi-) automated triangulation. Semi-automated triangu-
lation seems to be fairly well advanced in at least two major
vendors' systems. This is not so surprising considering the
fact that already in 1987 the DCCS had offered an
apparently operational software (Helava, 1987).
(a) Early investigations (Ackermann, Schneider, 1986)
have shown that the results of digital triangulation are
of the same accuracy as those of the triangulation with
analytical plotters. Considering the latest findings on
the very high accuracy of image measurement of
signalized, well-defined points, the digital triangulation
results should be even significantly better. The reason
that this could not be confirmed in empirical tests so
far has its origin primarily in errors introduced through
image scanning. Recently we have shown that with
direct digital image acquisition using a Kodak DCS 200
still video camera, the planimetric accuracy can be
improved by roughly a factor two to 1 micron in image
space (Kersten, 1996). However, with 0.1 %o flying
height the height accuracy is by a factor three well
below the performance of an aerial photographic
camera. This is partly due to the narrow bundle of rays
of the video camera (18 mm camera constant at a
CCD chip format of 14 mm x 9.3 mm).
At this point it seems worthless to conduct, accuracy
tests in digital triangulation based on natural object
points and on significantly distorted images by
scanning. Scanning errors can be better checked and
isolated otherwise (Baltsavias, Waegli, 1996), without
going through a full block triangulation procedure.
Furthermore, for the test of a system's accuracy
capability, signalized well-defined points should be
used.
(b) Another problem worth mentioning is that of an
appropriate pixel size for triangulation. If triangulation
is done with only natural points (control points, tie
points, new points) requirements concerning pixel size
are fairly relaxed and depend on the type of object
point and on the image scale. A pixel size between 20
and 40 microns should usually suffice. If, for high
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996
