the strength of multi-station geometry in aerial applica-
tions.
As the computational effort connected with the strict solu-
tion of ambiguities as published in (Maas, 1992a) grows
exponentially with the number of images, a reduced
version had to be implemented. While the strict solution
with three or four images is based on a combinatorics
algorithm, the reduced solution is based on probability
measures for potential correspondences, which are
defined by the number of images the candidate can be
traced through. This is implemented in a recursive manner
in a way that the longest traces (i.e. the points which can
be matched successfully in the largest number of images)
are accepted first and in case of ambiguities only a candi-
date with a trace, which is significantly longer than the
traces of the other candidates, is accepted. The flow chart
scheme in Figure 1 elucidates this principle.
1. take a candidate in image I; , 2. compute the Spinola} line
\ into an image I,, find candidate(s)
in the epipolar search area
3. verify match(es)
in all other images
lun e ou o EE EEE EE EE WA WA WA WA Wa WA wa wa wa wa
4. count number 5. accept trace
of successful verifications
i
> i
candidate 1: n; matches \ max (ny, iy, ..., fp)
1
1
|
significant
candidate 2: n, matches
-> match found
candidate p: ny matches
Figure 1: Computation scheme for the automatic establishment
of correspondences via epipolar line intersection
For a point P" in the first image the epipolar line in the
second image is computed. Then, for all candidates on this
epipolar line, the assumption of a correct match is verified
or neglected by intersections with the epipolar line of P' in
all other images. Finally, the number of successful verifi-
cations is counted and the match is accepted if the number
of successful matches for a candidate is large enough and
significantly larger than the number for all other candi-
dates. This procedure does not require all points to be
imaged or detected in all images. Moreover, traces do not
necessarily have to begin in the first image. In this
manner, all candidates in the epipolar search area are
486
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
either confirmed as valid candidates or rejected as false or
spurious matches.
The technique does not necessarily require base compo-
nents forming a triangle. Using a parametrization on the
epipolar line, it can also be applied to image data acquired
from multiple camera positions on a straight line, i.e. with
parallel base components and parallel epipolar lines
(Maas, 19922).
A similar approach is also used by (Lotz/Froeschle, 1990)
and a number of other authors, who combine short and
long base components to warrant both reliability and
precision by reducing the probability of the occurrence of
ambiguities in the establishment of correspondences
between two images. However, this approach does not try
to use the multi-image geometry to solve occuring ambi-
guities. It can however be shown (Maas, 1992a) that if
occuring ambiguities are solved the ideal baselength ratio
between three collinear camera stations for achieving a
maximum rate of successfully established matches is
bo: ba: biz=1 d 2,2;
A procedure for the solution of ambiguities has also been
published under the title multiple-baseline stereo by
(Okutomi/Kanade, 1993), who treat the trade-off between
precision and reliability in matching with an interesting
closed solution avoiding search procedures, which is
however limited to one-dimensional area-based matching
by minimizing the sum of squared differences.
PRACTICAL RESULTS
Besides to a number of applications in close-range photo-
grammetry, the n-ocular extension of the epipolar line
intersection technique has been applied to the derivation
of digital elevation models on two datasets:
e Dataset 1 consists of six scanned aerial images of the
Simplon pass area in Switzerland with 8090/6090
overlap.
e Dataset 2 consists of a block of 50 images taken from a
helicopter with a high resolution digital stillvideo
camera.
Dataset 1 - scanned aerial images
Images of a region on the Simplon pass in the Swiss Alps,
which shows elevations between 1400m and 2400m, were
taken from a flying height of 4600m, using a 150mm lens.
The nominal overlap in flight direction was 80% and the
overlap between stripes 60%, with relatively large devia-
tions due to terrain height changes. The black-and-white
images were scanned at 600dpi resolution (42 micron
pixel spacing, image size ~ 5200 x 5200 pixel) on an
uncalibrated Agfa Horizon desktop scanner. The fiducial
marks in the scanned images were measured with least-
squares-matching. A og of the affine transformation for
inner orientation of 60-70 micron indicates the limited
accuracy of the scanner.
For the DEM generation with the epipolar line intersec-
Figure
tion t
of th
form:
data
whole
Discr
sub-b
Fig
With
imag
the si
whol
flight
the ir
line i
of al
with
that
With
yield
