wu p= @~N =
Mw
Jorge Brito
Figure 7. Criteria for choosing the probabilities
Pr(a;) and Pr (bj).
3 TESTS
The preliminary step for executing of the tests for occlusion detection was the selection and measurement of ground
coordinates of point pairs in problem areas in the images. This task was performed in the digital photogrammetric
workstation DPW 770. The same pair of aerial, digitized conjugate images at nominal scale of 1:12,500, over the region
of Escondido, California, was used as the primary source of digital photogrammetric data.
The measurement of ground coordinates of testing point pairs in problem areas found in the images required that the
ground coordinates of the testing points be manually measured in the stereo-model. They could not be automatically
measured because of failures in automatic image matching procedures for finding conjugate point pairs in a digital
image. This restriction stresses the importance of the investigation being developed.
A total of 26 point pairs were selected in the stereo-model, mainly along building corners. Besides the measurements in
the stereo-model, it was also necessary to measure the same point pairs in the digital orthoimage in order to compare
results. To do so, it was necessary to compute the height (Z) of the testing points from the DEM used to produce the
orthoimage because only the planimetric ground coordinates (X, Y) can be obtained from the orthoimage.
A third data set was also generated. It has as its initial input the same ground coordinates as the stereo-model
measurements. The difference between them is explained by including the imprecision of the DEM (Opzm) used for
producing the digital orthoimage. Artificial errors were introduced in the projections of the stereo-coordinates onto the
digital image space as a consequence of this action. This fact justifies why the name “simulated” was assigned to this
data set. This data set was produced because of the deterministic nature of the results derived from the stereo-
measurements. This problem will be explained later in the discussion of the results. After measuring the ground
coordinates, it was necessary to compute the projections of the testing points onto the digital image space and their
precision through their respective variance-covariance matrices. This task was executed by the mathematical model for
error propagation through the photogrammetric processing.
4 DISCUSSION
The results of occlusion detection through a Bayesian network, for each of the data sets tested, were compared
together. This comparison is represented by the Venn Diagram shown in figure 8:
The following conclusions can be drawn from figure 9:
(1) All the approaches agree in their respective analysis of occlusions for 15 out of 26 pairs, or 57.7% of the time;
(2) the stereo and the ortho image results agree in the same proportion as computed above;
(3) the simulated data and the orthoimage agree in 17 out of 26 pairs, or 65.4% of the time and;
(4) the simulated data and the stereo-model have produced the same analysis of results for 24 out of 26 pairs, or 92.3 %
of the time.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 107