In order to become operational for practical applications
an automatic orientation module has to meet various
requirements. Ideally it should be
- autonomous (as opposed to the term 'automatic', 'au-
tonomous' implies that no user interaction whatsoever
is acceptable),
- faster than manual image orientation,
- more accurate than manual image orientation,
- flexible with respect to image and camera type (close
range, aerial, satellite images),
- flexible with respect to different types of control infor-
mation (points, lines, areas, DTM, etc.),
- robust: the module should also work with images of
poor quality,
- reliable: the module should have a self diagnosis pro-
cedure by which success and failure of the computa-
tions can be assessed.
This papers reviews existing algorithms, strategies, and
systems for automatic image orientation in digital photo-
grammetry. What is new in the digital domain is the
automatic extraction and matching of the image primiti-
ves. The actual computation of the orientation parame-
ters is similar, if not identical to the corresponding step
in analytical photogrammetry. Therefore, the emphasise
of this paper is on the automatic primitive extraction and
matching. The discussion is focused on aerial applica-
tions. In the next chapter some background will be given
on the central task of image matching. Then, interior and
exterior orientation are discussed, the latter split up into
relative and absolute orientation. Throughout the discus-
sions, examples from the literature illustrate the state of
the art in the field. The paper concludes with some
remarks on the future of automatic image orientation.
2 MATCHING FOR IMAGE ORIENTATION
Matching plays a basic role in the automation of image
orientation. Therefore, some background on this impor-
tant topic will be given in this chapter (see also Heipke
1996).
In digital photogrammetry and remote sensing, matching
can be defined as the establishment of the corre-
spondence between various data sets. The matching
problem is also referred to as the correspondence prob-
lem. The data sets can represent images, but also maps,
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
or object models and GIS data. Many steps of the photo-
grammetric processing chain are linked to matching in
one way or another. Examples include
- the reestablishment of the interior orientation: the
image of a fiducial is matched with a two-dimensional
model of the fiducial,
- relative orientation and point transfer in aerial trian-
gulation: parts of one image are matched with parts of
other images in order to generate tie points,
- absolute orientation: parts of the image are matched
with a description of control features
- generation of DTM: parts of an image are matched
with parts of another image in order to generate a
three-dimensional object description,
- the interpretation step: features extracted from the
image are matched with object models in order to
identify and localize the depicted scene objects.
Except in the case of interior orientation, in image mat-
ching we try to reconstruct three-dimensional object in-
formation from two-dimensional projections. During im-
age acquisition information was lost. This is most evident
in the case of occlusions. Image matching belongs to the
class of so-called inverse problems, which are known to
be ill-posed. A problem is ill-posed, if no guarantee can
be given that a solution exists, is unique, and/or is stable
with respect to small variations in the input data. Image
matching is ill-posed for various reasons. For instance,
for a given point in one image, a corresponding point may
not exist due to occlusion, there may be more than one
possible match due to repetitive patterns or a semi-trans-
parent object surface, and the solution may be unstable
with respect to noise due to poor texture.
In order to find a solution of an ill-posed problem one
usually has to deal with an optimisation function exhibi-
ting many local extrema and thus a small pull-in range.
Therefore, stringent requirements may exist for initial
values of the unknown parameters to be determined.
Moreover, usually there is a large search space for these
parameters, and numerical instabilities may arise during
the computations. Ill-posed problems can be converted
to well-posed problems by introducing additional
knowledge. Fortunately, a whole range of assumptions
usually holds true when dealing with photogrammetric
imagery:
- information about the sensor, e.g. in form of a calibra-
tion protocol is available,
-iithe;
red
banc
-"Sthe
fect
ima
= the
def
rigi
- the
zs the
- the
flec
= doit
the
Depen
assum]
list m:
which
rithm
rent al
Most
implic
tions €
In ord
decom
1994a
= the
Pos
ma
fea
des
chi
= the
me
im:
per
jec
nes
on
- the
ren
Fo:
inc
of
