| processing
routine. The
‘the village,
ation and
we shown in
contour line
idings, trees,
by the tech-
d points are
bjects above
nterest oper-
ability of the
all involved
1 only rather
) points of a
oot of build-
ing foot points
A thorough visual check of the digital orthophoto mosaic
with an overlay of matched points and contour lines
showed a total of only 12 gross errors in the dataset,
which corresponds to an error rate of 0.2%. Sharpening
the requirement for the acceptance of points to matches in
at least 6 images, 4270 object points with 5 gross errors
remained, and if points were only accepted when matched
in at least 7 images, the number of reconstructed object
points drops to 2750 (see Table 1).
Table 1: Interest operator efficiency
min. images 5 6 7 8 9 10
object points | 6866 | 4272 | 2757 | 1787 | 1187 | 757
In combination with the average overlap of ~70% (theo-
retical average visibility of points in 12.25 images) this
indicates a probability of approximately 65% for the Foer-
stner-operator detecting identical points in two images
and a probability of approximately 0. 651) for detecting
identical points in n images.
CONCLUSION
In contrast to most stereo-based techniques, the presented
technique does not avoid ambiguities in the establishment
of correspondences by image pyramids, but solve them by
a consequent exploitation of the geometric strength of
multiple images, implemented via the intersection of
epipolar lines or multi-baseline techniques. The method
shown in the presentation can be characterized by the
following keywords:
* High reliability due to the multi-image geometry.
* No approximate values are needed, no interactively set
start points, no image pyramids are involved.
* No a priori knowledge on maximum terrain slopes is
required, the terrain slope can be practically unlimited.
* Matching of discrete points avoids smoothing effects
connected with many area-based DTM generation
methods.
* Good precision due to subpixel interest-operator and
measurement in multiple images.
As far as the allowable arbitrary terrain discontinuities are
concerned, these options exceed the requirements posed
by the automatic generation of digital elevation models in
most terrain types. However, the aim of the study to show
the potential of the consequent use of the strength of
multi-image geometry for the robust establishment of
correspondences, could be achieved. Blunder rates of less
than 0.5% of the matches without the use of pyramid
approaches or given approximate values can be consid-
ered a very good result for automatic DEM generation in
mountain regions.
One limitation of the technique is the requirement that
points have to be imaged in multiple images, which neces-
sitates 80%/60% image overlap in mountain regions. A
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
major limitation is the probability of interest operators
detecting identical points in multiple images. In the image
material used in the presented study, the probability that a
point, which was detected by the Foerstner operator in one
image, was also detected in a second image, was found to
be approximately 65%, reducing the number of successful
matches exponentially with the required number of
images. Although possibly not dense enough for a good
terrain description, the presented results may be very
useful as approximate values for other techniques.
References
* Ackermann, F., 1994: Digital elevation models - tech-
niques and application, quality standards, develop-
“ ment. IAPRS Vol. 30, Part IV, pp. 421-432
e Baltsavias, M., Stallmann, D., 1992: Advancement in
matching of SPOT images by integration of sensor
geometry and treatment of radiometric differences.
IAPRS Vol. XXIX, Part B4, pp. 916-924
* Dold, J., Maas, H.-G., 1994: An application of
epipolar line intersection in a hybrid close range
photogrammetric system. IAPRS Vol. 30, Part V
e Fórstner, W., 1986: A Feature Based Correspondence
Algorithm for Image Matching. IAPRS, Vol. 26, Part
3/3, pp. 150-166
* Grün, A., 1985: Adaptive least squares correlation: A
powerful image matching technique. South African
Journal of Photogrammetry, Remote Sensing and
Cartography 14 (3), pp. 175-187
e Kersten, Th., 1996: Aerotriangulation mit einer digi-
talen Still Video Kamera Kodak DCS200. VPK 2/96
e Krzystek, P., 1991: Fully automatic measurement of
digital terrain models. Proc. of the 43. Photogram-
metric Week, Stuttgart, Germany, pp. 203-214
* Lotz, R., Fróschle, E., 1990: 3D-Vision mittels Stereo-
bildauswertung bei Videobildraten. 12. DAGM-
Symposium Mustererkennung, Informatik Fachber-
ichte 254, Springer Verlag
e Maas, H.-G., 1991: Digital Photogrammetry for Deter-
mination of Tracer Particle Coordinates in Turbulent
Flow Research. Photogrammetric Engineering &
Remote Sensing Vol. 57, No. 12, pp. 1593-1597
e Maas, H.-G., 1992a: Complexity analysis for the deter-
mination of image correspondences in dense spatial
target fields. IAPRS, Vol. XXIX, Part B5, pp. 482-485
e Maas, H.-G., 1992b: Robust Automatic Surface
Reconstruction with Structured Light. IAPRS, Vol.
XXIX, Part B5, pp. 102-107
e Okutomi, M., Knade, T., 1993: A Multiple-Baseline
Stereo. IEEE transactions on Pattern Analysis and
Machine Intelligence, Vol. 15, No. 4, pp. 353-362
e Schewe, H., 1987: Automatic Photogrammetric Car-
Body Measurement. Proc. of the 41. Photogrammetric
Week, Stuttgart, Germany, pp. 47-57
e Ursem, R., 1994: Accurate Reconstruction of a 3D
Wireframe of a Human Head. Diploma Thesis TU
Delft, The Netherlands
