nts
als described
features
als managed
al database
als possible,
neg. determi-
n
lucials only
to missing
Istment was
he mapping
icy of aerial
e problems
the exterior
ial triangu-
reo plotting
1cluded that
orientation
'estigated is
act, various
]ready exist
95; see also
er by Först-
y Haumann
ble systems
he different
identifying
determining
images can
ive and the
ile AAT is
nt in image
ve and the
which need
issues such
/ith varying
olumes etc.
| in Schenk
ntations of
of indirect-
ly determining the corresponding parameters from ima-
gery, a few thoughts about the relevance of this task in the
era of GPS and INS and sensor integration are in order.
In theory, GPS and INS allow for the direct measurement
of the exterior orientation parameters and thus render
photogrammetric solutions for this task obsolete. The
main issue is that ground control as such is not needed in
the scenario of GPS/INS photogrammetry, and thus the
resulting multi sensor data acquisition device becomes
totally autonomous (see e.g. Ackermann 1995b for a
discussion on the possibilities of autonomous multi sen-
sor systems). The accuracy requirements for the orienta-
tion parameters of various photogrammetric applica-
tions are discussed and compared to available GPS and
INS measurement accuracies e.g. by Schade (1994) and
Schwarz et al. (1994). Among others Ackermann (1994),
Burman, Torlegärd (1994) and Hothem (1995) report on
the state of the art of aerial triangulation using GPS
observations for the projection centres of the camera.
Without going into detail it is concluded here that while
the impact of GPS and INS on photogrammetric orien-
tation is already large and still growing, photogrammetry
without ground control is not yet a reality. However, any
automation in image based exterior orientation proce-
dures has to be seen and judged in the light of the
developments in the direct measurement of the exterior
orientation parameters using GPS and INS.
4.1 Automatic relative orientation
The relative orientation of two overlapping images des-
cribes the relative position and attitude of two images
with respect to one another. It is a 5-parameter problem.
Given these 5 parameters all imaging rays of conjugate
features intersect, and these intersections form the model
surface. After having completed the interior orientation
for both images separately, the two image coordinate
systems are explicitly known. Therefore, relative orienta-
tion is a non-semantic task, and arbitrary conjugate fea-
tures can be used for the computation of the orientation
parameters. It must only be assured that enough features
distributed across the complete model are used.
A general, autonomous module for relative orientation
should be fast, accurate, robust, and reliable (see again
chapter 1). Further it should not require any approximate
values (in particular scale and rotation invariance should
be available), and the approach should ideally work with
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
multi temporal, multi spectral, and multi sensor imagery.
The input should only consist of the images themselves
and the results of interior orientation, the output are the
five orientation parameters, the three-dimensional coor-
dinates of the conjugate features, and corresponding
accuracy measures.
A generic solution for autonomous relative orientation
involves the following steps:
- compute image pyramids for both images separately,
- approximately determine overlap and possible rota-
tion and scale differences between the images on the
highest level,
- extract features, possibly including relations,
- match these features (and relations),
- determine coarse orientation parameters,
- proceed with extraction, matching, and parameter de-
termination through the pyramid from coarse to fine
in order to increase the accuracy of the results.
Image pyramids should be employed to take advantage
of the concept of hierarchy already mentioned. Note that
the repetition of feature extraction, matching, and para-
meter determination from one pyramid level to the next
leads to a close integration of the coordinate measure-
ment and the actual computations, two tasks which are
well separated in analytical photogrammetry. In view of
what was discussed about image matching in chapter 2
the mentioned steps will be described in more detail.
In order to detect overlap, rotation, and scale differences
between the images matching primitives which are inde-
pendent on absolute position, rotation and scale must be
used. The cross correlation coefficient is known to be
neither scale nor rotation invariant. Least squares mat-
ching can't be used either, because it requires accurate
approximate values for the unknowns which are not avai-
lable. Thus, area based matching is not an appropriate
method for this task. Feature based methods can be
employed to detect rotation differences between images.
For example, straight lines can be detected in both im-
ages, followed by a comparison of the histograms of line
direction. A detection of scale differences on the basis of
the line length, however, is more problematic, because
lines are often broken up into small pieces. Note that the
same argument motivated the design of the line extrac-
tion algorithm by Burns et al. (1986) now widely used in
computer vision and photogrammetry. Rotation invari-
303
