THREE DIMENSIONAL OBJECT RECONSTRUCTION BY OBJECT SPACE MATCHING
Franz Rottensteiner
Institute of Photogrammetry and Remote Sensing
TU Vienna
GuBhausstraDe 27-29
A-1040 Vienna, Austria
frottens @fbgeo1.tuwien.ac.at
ISPRS Commission Ill, Working Group 3
KEY WORDS: Matching, Surface, Modeling, Reconstruction, Three-dimensional Object Reconstruction,
Feature Adjacency Graphs, Object Space Matching
ABSTRACT:
One of the main current fields of research in photogrammetry is concerned with the reconstruction of object surfaces
from digital images. Both feature based and area based matching algorithms have been described in literature. Most of
these algorithms aim at a 2:4D representation of the object surface which is to be derived from a digital stereo pair.
They give good results when the surface to be reconstructed is smooth, and they face problems in areas with surface
: discontinuities. Feature based matching techniques appear to be more flexible with respect to surface discontinuities
and requirements for approximate values; area based matching techniques offer a higher accuracy potential. In the
course of the Austrian Research Program on Digital Image Processing we have developed a concept for an image
matching algorithm considering topological relations which should work under quite general circumstances, e.g. in
close range applications. A generalized framework based on the Fórstner operator will be used for feature extraction.
The generation of correspondence hypotheses will be based on region adjacency graphs; we are investigating more
complex models in object space for the evaluation of these correspondence hypotheses in order overcome the problem
of occlusions. The integration of a bundle block software should provide geometrical constraints for correspondence
hypotheses and give us the possibility to use more than two images for surface reconstruction. Approximate values will
be improved by hierarchical methods (image pyramids) starting from a very coarse level. The final result of our
algorithm should be a fully three-dimensional representation of the surface to be reconstructed.
1. MOTIVATION
One of the main current fields of research in photo-
grammetry is concerned with the reconstruction of object
surfaces from digital images. A high level of automation
can be achieved for this task by applying matching
techniques to digital images (Gülch, 1994). Several
different approaches have been proposed in literature.
Some systems have left the experimental state and can
now be used in the production process of Digital
Elevation Models (DEM), at least for small and medium
image scales. These systems aim at a 25D representa-
tion of the object surface which is to be derived from a
pair of epipolar images (Krzystek, 1995); (Gülch, 1994).
The matching algorithms used in such systems give good
results when the surface to be reconstructed is smooth
but face problems in areas with surface discontinuities,
especially in urban areas, because most of them
implicitly use models assuming the object surface to be
smooth or even planar. lf the object meets this
smoothness assumption and if it can be modeled in 25D,
the above algorithms can also be used for close-range
applications. However, if those conditions are hurt, they
will fail to give good results.
In the course of the Austrian Research Program on
Digital Image Processing we are investigating matching
techniques for object reconstruction which can also be
used in cases when the requirements mentioned above
are hurt, as it happens with many close range
applications and with large scale aerial images. Less
emphasis is laid on the optimization of computational
speed than on finding matching algorithms which give
accurate and reliable results under quite general
circumstances and render possible a high degree of
automation. The demand for applicability of the algorithm
to close-range images and for large aerial image scales
leads to certain requirements:
« Many classes of objects can no longer be described
by 21D models. Thus, more sophisticated methods
for fully 3D representation of the object to be
reconstructed become necessary.
e |n close-range photogrammetry, more than two
images of a certain object region are usually
available. The viewing directions might be convergent,
thus the assumption of a near-normal-case con-
figuration might have to be dropped.
e Matching will be heavily influenced by surface
discontinuities and occlusions; the object surface can
thus no longer be modeled to be smooth.
e In order to achieve a high degree of automation, very
coarse approximations for the object surface should
be sufficient for the algorithm.
In section 2 we want to give an overview about related
work in the fields of image matching and object
representation. In that section we will also discuss the
applicability of the techniques described in literature to
our problem. Our concept will then be described in
section 3.
- This work is funded by the Austrian Research Program S7004-MAT on Theory and Applications of Digital Image
Processing and Pattern Recognition, Project IV: Stereovideometry and Spatial Object Recognition
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
692
211
One
give!
for tl
Fi
Matc
reco
pers
foun
imac
the «
that
Soph
the
num
proc
the «
the €
appl;
(Krz
map
impli
be a
surfe
assu
is the
terra
occlt
Fron
char:
1994
repre
map
com
base
a sy
featu
oper
then
assu
be
1995
flexit
requ
matc
(Gül
Both
in tl
distri
whic
