The following‘ observations can be made from the
obtained results of the successful ARO runs:
- The models checked on the PHODIS ST stereoplotter
were all found to be free of y-parallaxes.
- The o, as a measure of the overall accuracy ranges
approximately between 0.2 and 0.4 pixel.
- 70-190 well distributed conjugate pairs provide a high
reliability for the five estimated orientation parameters.
- The elapsed computing time amounts to 75 sec (30um
pixel size) and 2-3 min (15pm pixel size) for an stereo
pair on a Silicon Graphics Indy with R4400 processor
(150 MHz).
Tables 3-6 show the o, for different image scales, terrain
types, pixel sizes, overlap and number of conjugate pairs.
It can be seen from Table 3 that the terrain type has no
significant influence on the accuracy of the results, and
that the accuracy is getting only slightly poorer as the
image scale decreases. The accuracy, however, is highly
dependent on the pixel size of the scanned images. Table
4 shows that o, for images with a large pixel size (mostly
30um) is about half of the value for images with a small
(mostly 15um) pixel size. This result is in contradiction to
the assumption that the accuracy of image matching
mainly depends on the pixel size. At this stage we have
no explanation for these findings, and further theoretical
and empirical investigations will be conducted to clarify
this point.
As represented in Table 5 an end overlap of 80% or more
provides a significantly better accuracy than 60 %. This
finding is not surprising, because the larger the overlap
the more similar are the images. The influence of the
number of conjugate points on O, is clearly visible in
Table 6. ©, decreases with increasing number of
conjugate points.
As a typical example the results of the stereo pair Lohja
are depicted in Figure 1l. The two images are
superimposed with the extracted conjugate points at level
0 of the image pyramid.
In three of the special cases ARO failed to produce
correct results. The reason is that these cases violate at
least one of the assumptions incorporated into the
algorithm. It should be noted, however, that these extreme
cases do not occur in usual aerial photogrammetry.
- Homburg: the image scale of the two images differs
too much,
- Istanbul: the overlap is too small,
- Burghausen: the rotation difference is too large.
The other three of the special cases from Table 2 could be
processed successfully. This was not much of a surprise
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
for the spaceborne example depicted in Figure 2.
Interesting is the fact that the problems with the
dangerous cylinder in the example Schelingen could be
overcome. The algorithm extracted less conjugate points
as compared to the other cases, and the accuracy was not
excellent either, but the subsequently computed epipolar
imagery was free of y-parallaxes. This is in contrast to an
unsuccessful attempt to manually orient the stereo pair on
an analytical plotter. Perhaps surprisingly, ARO also
performed well for the close range example Felsendom.
The results are shown in Figure 3. This example should,
however, not be interpreted as a statement, according to
which ARO can handel close range imagery. More
experiments need to be performed in order to assess the
potential of our algorithm for such applications.
5. CONCLUSIONS AND OUTLOOK
The paper reports on the results of an investigation of
automatic relative orientation. Concept, algorithm and
realization are described. Test strategies and runs with 53
different stereo pairs are presented. About 100 well
distributed point pairs were selected for each model,
except for three extreme cases. Many more point pairs
could be made available. The obtained root-mean-square
standard deviations of image coordinates generally lie
between 0.2 and 0.4 pixel. Stereo models were found to
be free of y-parallaxes by skilled human operators. The
elapsed computing time was about 75 sec per stereopair
scanned with 30 um pixel size and in between 2 and 3
min for a 15 um stereo pair on a Silicon Graphics Indy
with R4400 processor (150 MHz). It could be proven that
the automatic relative orientation procedure is ready for
photogrammetric practice.
6. REFERENCES
Ackermann, F., 1983. High Precision Digital Image
Correlation. Schriftenreihe des Instituts für
Photogrammetrie der Universität Stuttgart, Heft 9, 231-
243.
Braun, J., L. Tang, R. Debitsch, 1996. PHODIS AT - An
Automated System for Aerotriangulation. Paper accepted
for ISPRS Congress'96, Vienna, Austria, July 9-19, 1996.
Dôrstel C., 1995: PHODIS innovations, in: Fritsch D.,
Hobbie D. (Eds.), Photogrammetric Week ’95, Wichmann,
Heidelberg, 5-10.
Haala, B., M. Hahn, D. Schmidt, 1993. Quality and
Performance Analysis of Automatic Relative Orientation.
Proceedings of the SPIE Conference on Integrating
Ta
Re
Te
Ke
Ta
the
the
Ph
Fi
of
