Page 97 - NGTU_paper_withoutVideo

Modern Geomatics Technologies and Applications

3. ESTIMATING THE ABILITY OF CLUSTERING METHODS TO SELECT CONTROL POINTS
In this step, to assess the precision of the clustering method in determining the position of the control points, the drone images
were oriented using the control points closest to the
cluster centres. Then, the residuals at the checkpoints
were used to evaluate the capability of the clustering
methods. In this paper, nine points around the study area
(Figure 4) were used as a checkpoint. Then, the
clustering methods were evaluated from three
perspectives: accuracy, stability, and distribution. In
the following, in addition to explaining how each criterion
is determined, the results of the relevant experiments are
also reported and discussed.

3.1 Evaluating the accuracy of clustering
methods to select optimal control points
Fig. 8. Comparing the accuracy of clustering techniques
In this section, we against the Grid methods using 60%-Overlap images examine the accuracy of
clustering algorithms in choosing optimal location of
the control points. For each algorithm, the clustering was
performed three times with the number of clusters 5, 9,
and 25. The reason of choosing different number of
clusters was to study of the ability of the algorithms in
selecting the appropriate positions with different
number of control points used. Anyway, after
clustering, the control points near the cluster centres were
determined and used to do the absolute orientation of the
images. In addition, to ensure Fig. 9. SOM results the accuracy of the clustering
results, the images were also oriented using 5, 9, and 22
points regularly distributed across the area of test. In our
test, this is referred to Grid.

Figure 6 shows the results for images having 90% overlap. In this figure, the vertical axis shows the Root mean square error
(RMSE) error of the check points (in cm) for each of the clustering methods (shown on the horizontal axis). As can be seen, the
relative relationship between the accuracy of clustering methods in different situations (i.e. different number of control points)
is almost the same. Besides, the errors of the Optic and Dbscan methods are more than those of the other techniques. Except
these two, the accuracy of the other methods is almost the same which is almost equal to that of the Grid method. A question
that arises here is whether the above results are only for cases where the overlap is 90% and the network geometry is strong, or
they could be achieved even with smaller overlaps. To examine this, two other series of images having smaller overlaps were
extracted from the 90%-overlap images and the accuracy of the clustering methods was investigated once again. Figures 7 and
8 show the results of this experiment.

Fig. 6. Comparing the accuracy of clustering techniques
against the Grid methods using 90%-Overlap images

5