Improving Geolocation by Combining GPS with Image Analysis - Geoinformatics for Intelligent Transportation - page 210

In this case, if the GPS information is accurate enough for the geolocation, then
the latitude and longitude coordinates may be used to calculate the distance from
the current position to the closest set of reference points. This calculation has the
purpose of discarding the more distant POIs, reducing the number of valid POIs for
image comparison, at any time. If, on the other hand, the received GPS information
is not accurate enough to be reliable, then the image analysis is started, to com-
pensate that lack of accurate geolocation information. If the system suddenly stops
receiving GPS information, a timer is started. This timer stops either because GPS
data is available again or because a maximum time Y is exceeded. In the second
situation the image analysis process is started. When this process takes over, if the
result of analysis is positive the system

knows that it is close to the geographic coordinates associated to the matched POI.

— a POI has been detected in a video frame —

2.2 The Image Analysis Component

The image analysis process requests images from two sources: an external camera,
placed near the GPS receiver, capturing frames in real-time (observed image) and a
database of POIs (model images). From the database, only the images of POIs
closer than a threshold X to the actual location are considered. In order to obtain
good performances, the number of image comparisons should be the lowest pos-
sible, avoiding unnecessary image analysis. To analyse the images, a feature
descriptor is used, which detects characteristic elements (features) between two
images and compares them trying to find similarities. In our approach, the SURF

(Speeded-up Robust Features) algorithm has been used, due to its scale invariance
property, which is an important factor to consider when capturing images at distinct
distances, affecting the scale of the point to detect [ 5 ]. The performance of SURF
has also been taken into account, compared to other feature descriptors [ 9 ]. The
necessary time to analyse each captured frame is an important factor to consider in
order to be able to process regions with a higher density of POIs. To guarantee
better results, it is important that every reference point has more than one associated
reference image, captured from distinct points of view. Ideally, for every POI the
database should hold at least three images, one frontal and two lateral, in which case
only one match would be necessary to obtain a positive result. This three images
allow us to contour the partial lack of rotation invariance of the SURF algorithm.
Figure 2 shows outputs generated by the image analysis process.
As it is possible to see in Fig. 2 a, the image analysis process found a match, but
clearly (red line shape depicts result from algorithm) it does not correspond to a
POI, generating a false positive situation. In order to provide better results and
discard situations like this, it is important to implement a

filter for False Positive
discarding. The false positive detection was based in the analysis of the red
quadrilateral (more speci cally of the four points returned by the algorithm). If the

points do not represent a quadrilateral (as depicted in Fig. 2 c), the system considers
it as a false positive and discards it, because in true positives, the shape resembles a