Improving Geolocation by Combining GPS with Image Analysis - Geoinformatics for Intelligent Transportation - page 217

satisfactory. In future tests, perhaps it would suf ce to use the image analysis within

a distance threshold of 100 150 m to the POI. Finally, regarding performance of the

system, concerning the image descriptor, although the results of the image analysis
are very good (considering the number of analysed images and the number of false
positives detected) the time that each analysis took might still be improved by using
a faster image descriptor. An example of a faster algorithm is ORB (Oriented
BRIEF) detector [ 10 ]. Theoretically, this new detector can perform more analysis
per second, keeping the same success rate than SURF detector.

-

7 Conclusion

This work presents an innovative approach for keeping geolocation accurate in
mobile systems that rely mostly on GPS, by using computer vision to help pro-
viding geolocation data when the GPS signal becomes temporarily low or even
unavailable. For some applications, for instance, a city tour-bus transportation that
shows location-sensitive data in its screens as it passes by a POI, the lack of GPS
data, even for a short period of time, may lead to undesired results.
The main contribution of this work is a method that enables geolocation by using
feature recognition from computer vision and GPS technology in a complementary
fashion. When available, GPS signal is used in order to know the distance from the
mobile system where the CV-GPS is assembled, to the nearer POIs. This way, it
prevents the CV system from trying to find a match with the entire set of POIs

available in the database. When GPS signal is unavailable the CV module (feature
recognition) is used to identity POIs in video frames, captured by a video camera
placed on a mobile system. If a POI is matched to one of the POIs available in the
database containing model images, then it is assumed the mobile transport is known
to be near GPS coordinates associated to the matched POI. As soon as GPS data
restarts being available, the computer vision system stands by.
To test our method we de

ned a set of experiments (appropriate variables and
test scenarios). The resulting set consisted of eight experiments, some of which
were repeated more than once. The experiments targeted different goals: to test the
simple detection of POIs, to test false positive situations by forcing these situations,
to simulate situations of lack of GPS signal and simulate urban canyon situations
(no GPS signal). Three distinct routes and three distinct POIs were used in our
experiments, with different characteristics, such as, different road conditions or
POIs texture or shape. The atmospheric conditions under which the tests were
conducted were not ideal (cloudy day), with less favorable light conditions.
The results achieved were positive for almost every experiment: only few false
positives were detected. Good results were achieved with different floor conditions,

vehicle velocities and with or without GPS information available, returning only
four bad result instances. The four false positive detections are acceptable if we
consider that 1,472 images were analyzed. Furthermore, we detected a pattern in
the identi

fl

ed false positives (the pattern recognition result returns a concave