Geoscience Reference
In-Depth Information
detected in the third experiment, and it was similar to the one detected in the
rst
experiment. The fourth experiment was performed over a bigger route (2.5 km), but
only 18 % of it was covered by image analysis. In this experiment it was possible to
verify that the GPS information was properly working together with the image
analysis, in order to minimize the detections in places where no POI was available
for detection. At the same time, it was possible to verify that no false positives were
detected in this route. The
fifth experiment had the deliberate objective of mislead
the system. By introducing incorrect information in the database it was possible to
verify if the image analysis component would mix up the POIs and detect a not
existing POI in the image (the POI did not even belonged to that route). This
experiment was repeated twice. In the
first repetition, 5.1, a false positive was
detected, but not where it was expected. The false positive was detected around
180 m of the coordinates of the target POI, which means that the system did not
confused the POI1 with POI2, but failed in a common detection as in the false
positives detected in the experiments number one and three. In the second repeti-
tion, 5.2, the experiment went as expected and no POI was detected. The sixth
experiment tried to simulate an urban canyon situation, where suddenly the GPS
coordinates stop being available. In order to simulate this, as explained in the
previous sections, the GPS was turned off which increased the percentage time of
image analysis of this route to 34 %, because the image analysis was started after
2 s without GPS input. The experiment returned positive results, and POI1 was
detected in time, without false positives. In the seventh experiment the goal was
once again to simulate an urban canyon situation by deactivating the GPS infor-
mation, but using a different route. The results were as expected and POI2 was
detected in time. The eighth and last experiment, evaluated a problematic situation
that could not be solved by a GPS only solution. A method using GPS information
only would consider that the vehicle was passing by the POI in a situation where the
POI would not be visible yet but with this system, the POI was detected in the
proper time, and no false positives were detected.
Summarizing, we veri
ed that the system performed as expected in most cases.
The light conditions in which the system was tested were not ideal at all, and
nevertheless the system responded quite well. With better atmospheric conditions, it
is expected that the results are at least this good, because the quality of the captured
images would be superior. The false positives detected in tests 1.2, 1.3, 3.2, and 5.1
occurred relatively far from the expected POIs, at variable distances from 180 to
220 m, and can be discarded by simply reducing the threshold distance used to
trigger the image analysis (which is clearly too high), or by
fixing the problem in
the image detection stage, by adding another
filter to the results obtained from the
image analysis algorithm. This
filter would discard the concave quadrilaterals
returned in the detection, pattern veri
ed in all the false positives detected. Fur-
thermore, the 300 m distance threshold used in the experiments proved themselves
to be an overkill, because in most cases the POIs were not even visible at that
distance. Nevertheless, by using that distance was possible to keep the image
analysis running for a longer period of time, allowing a better study of the false
positive detections and guaranteeing that the number of false positives was already
Search WWH ::
Custom Search