urement of
an interest-
nsidered an
hnique, and
niques, the
of multiple
tion yet, the
area-based)
ffects.
of a regular
der rates of
essing.
ramid; this
commercial
ystek, 1991)
nn, 1992).
ts one (or a
ie matching
the starting
'rrain slope.
e industrial
F (Schewe,
it may show
ients, occlu-
lage texture.
)s in image
approximate
he consecu-
not require
ng approxi-
.is indepen-
lues or pre-
knowledge on the maximum terrain slopes; only very
rough pre-knowledge (like estimated minimum and
maximum terrain height for the whole observation region)
is required. The procedure is similar to the procedure
followed by MATCH-T: Discrete points in the images are
extracted by an interest operator (Foerstner, 1986), and
correspondences between points are established using
epipolar lines. These epipolar lines may become rather
long when the terrain shows large height differences,
leading to multiple candidates on the epipolar lines and
ambiguities in the establishment of correspondences,
which can often not be solved. For that reason MATCH-T
uses image pyramids as a coarse-to-fine method, thus
limiting the length of the epipolar lines at each level of the
pyramid. Furthermore, the ‘general overkill philosophy’
of MATCH-T (Ackermann, 1994) solves some of the
problems connected with DEM data acquisition algorith-
micly: By switching from interpolation to adjustment,
outliers generated by false matches or by objects above
the terrain surface (like single trees) can be removed by
robust surface fitting, and it has been shown that even
breaklines can often be detected automatically in the
dense datasets of typically more than 500'000 surface
points per stereo pair.
In the method presented in this paper, the length of the
epipolar lines is basically unrestricted, and ambiguities
due to multiple candidates on the epipolar lines are solved
by epipolar line intersection techniques in multiple
images, thus avoiding any smoothing effects introduced
by surface fitting techniques or by patch sizes in area
based techniques.
EPIPOLAR LINE INTERSECTION
The concept of the epipolar line for the establishment of
image correspondences is well known in photogram-
metry: If the orientation and camera calibration parame-
ters of images are known, for each point P' in one image
an epipolar line in an other image can be defined on which
the corresponding point P" has to be found. The length of
this line can be restricted if approximate knowledge about
the depth range in object space is available. Adding a
certain tolerance width to this epipolar line segment (due
to data quality and measurement errors) the search area
for the corresponding point location in the other image
becomes a narrow two-dimensional window. Depending
on the number of detected points per image, the arrange-
ment of points and the complexity and depth extension of
the object a problem of ambiguities will occur here, as
often more than one candidate will be found in the search
area. There is a general trade-off in stereo matching
between the requirements of precision (requiring long
epipolar lines and thus also long baselines) and the
requirements of reliability (requiring short epipolar lines
and thus short baselines to reduce the number of ambigu-
ities). If additional point features do not allow for a reli-
able distinction of candidates, these ambiguities can often
not be solved by a system based on only two cameras. It
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
has been shown in that the probability of such ambiguities
grows linearly with the depth extension in object space,
linearly with the width of the search area and with the
square of the number of points (Maas, 19922). A solution
of this problem based on the intersection of epipolar lines
in an image triplet has been shown in (Maas, 1991), where
trinocular correspondences in digital images of some 1000
particles marking a turbulent flow were established. The
same technique has been applied in combination with a
projected dot raster for surface determination in (Maas,
1992b). Also an extension to a four-camera system is
shown in (Maas, 19922). It can be shown that a reasonably
chosen additional camera station reduces the probability
of ambiguities by a factor of 5 to 100 - depending on the
number of points, object geometry and data quality.
The method has meanwhile been extended from trinocular
vision to n-ocular vision to be used with an arbitrary
number of image coordinate datasets, where the data (=
image coordinates of discrete points) may origin form
digital, analog or hybrid systems in close range applica-
tions as well as in aerial photogrammetry. The aims of this
development were:
* The method should be applicable for the establishment
of image correspondences between image coordinate
datasets from an arbitrary number of images (with the
additional requirement that the images have been taken
from at least three different camera stations).
* Features which would allow for a distinction of points
are not needed (but can optionally be employed).
* Only minimum knowledge about the object space is
required. This knowledge can be reduced to rough
approximations of the volume boundaries in object
space; approximate values for the features are not
required, the features may be arbitrarily distributed
over the object space. A continuous object surface is
not required, or the object surface may show an arbi-
trary complexity; this includes objects freely distrib-
uted in space (like particles in water), occlusions or
targets fixed on wires.
* The method is robust against missing points: Not all
points have to be measured in all images, points may
be missing due to occlusions, failures in the detection
procedure, illumination effects, limitation of the sensor
format, etc.
The tool was for example used for deformation measure-
ments on a masonry wall using 18 images (Dold/Maas,
1994) and for the modeling of a human head from 40
images (Ursem, 1994).
The aims of the development were mainly derived from
tasks occuring in close range photogrammetry, and they
exceed the problems occuring in automatic DEM genera-
tion, especially concerning the allowance of arbitrary
discontinuities in the object. Nevertheless, the application
of the technique to the generation of DEMs promises
some advantages over conventional methods, justifying
research on the benefits of a consequent exploitation of
485
