International Archives of the Photogrammetry,
reason for the rejection of a lot of IBPCO features in areas of
low or repeatedly texture (fig.9).
a) True correspondences b) Accepted correspondences
Figure 7: A selection of the corresponding features
In fig. 7 a) a selection of the true correspondence matrix 1s
visualized which is calculated from the known transformation
parameters between the point sets. The mentioned
neighborhood of the features is evident within the cluster in the
matrix. Obviously, in the upper left and lower right part no
correspondences exist, because the used cloud patches do not
cover the same scene completely. Fig 7 b) shows the result of
the accepted correspondences. They are located in a preferred
area where the viewing angle is almost collinear to the normal
of the discrete object.
b) False accepted
correspondence
a) True correspondence
Figure 8: Discrete orthoimages of true and false corresponding
candidates
Figure 9: Distribution of matched correspondences
However, more than 80% of the accepted 16 correspondences
conform to the true candidates. In fig. 8 an example for a true
and a false accepted correspondence is provided. Because of the
Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
repeated texture, the candidates can not be clearly distinguished.
In contrary, larger grids of the discrete orthoimages would
reduce this problem. Fig. 9 visualizes the distribution of the
accepted correspondences. The correspondences are highlighted
with a red circle. Further all extracted features of the IBPCO
process in the used scene are highlighted with yellow triangles.
The falsely accepted correspondences are highlighted with a red
Cross.
It has been illustrated, that possible corresponding candidates
(yellow triangles) are located in the whole scene. At the same
time only some regions show sufficient texture patterns for
successful cross correlation. In addition to the possible
enlargement of the discrete orthoimages, further attributes are
needed to differentiate the candidates. For demonstration, in fig.
10 the resulting value of the SUSAN operator is used to refine
the matching strategy. The figure shows the candidates of the
highest similarity. The goal is to find the adequate weights
between all included criteria. Further investigations have to be
made for such weighting strategy.
Figure 10: Corresponding features matched with the resulting
SUSAN value
5. CONCLUSIONS
In this research the strength of the combination of laser range
devices and photogrammetric images is shown for registration
purposes. An operator for feature extraction is developed based
on experience in digital image processing and point cloud
registration. The concept is introduced and validated with a
selection of the clouds of two view points. The results of the
accepted correspondences are analyzed. False correspondences
occur in cases of ambiguous texture. It is explained how to
reduce such cases by using larger grids for the cross correlation.
For the last stage, the calculation of transformation parameters
from the accepted correspondences a robust method is needed.
Therefore RANSAC (random sample consensus) published by
Fishler and Bolles (1981) is a good strategy to detect the
blunders. Thus, a success of 80% of true candidates is enough
for a reliable registration. More important is the distribution of
the candidates, which should be supervised by the
neighborhood or topology of the candidates.
Further investigations will be necessary to analyze the presented
operator in more detail. Especially the thresholds according to
the locally geometric situation should be controlled
automatically in an intelligent way. Also different investigations
will be made to judge the texture of the discrete orthoimage,
e.g. with the Haralick parameters, (cf. Luhmann, 2000), that
have to implemented in the IBPCO.
International Archive
MC tren dr CHER
Credits go to Claus
and geoinformatics
for providing the dat:
Besl, P.J., Mc Kay, I
D shapes. In: [EEL
Machine Intelligence
Chen, C-S., Hung, °
DARCES: A New A
Partially Overlapping
Pattern Analysis ane
1229-1234.
Fischler, M., Bolles,
paradigm for model
and automated carto
381-395.
Gühring, J., 2001. R.
and validation using
Conference on Rece
Modelling, Quebec C
Gühring, J., 2002.
mittels Streifenprojek
560, München.
Heipke, C., 1992: A
matching and surfac
(58) 3, pp. 317-323.
Huber, D.F., Hebert,
multiple 3D data sets
Computer Vision Bey
Johnson, A., Kang,
Textured 3-D Data.
Advances in 3-D Digi
Lavallee, S., Szelisk
orientation of free-fo
distance maps. In: /£
Machine Intelligence.
Lowe, D.G., 1999. OI
features. Internation.
1150-1157, Corfu, Gi
Luck, J., Little, C., I
using a hybrid simul
algorithm. Proc. Of I
and Automation, San
Luhmann, Th., 200
Wichmann Verlag, H
Neugebauer, P.J., 19
via simultaneous re
multiple range imag
Modeling, Vol. 3 (1&