Improving Geolocation by Combining GPS with Image Analysis - Geoinformatics for Intelligent Transportation - page 208

Besides this fact, both SIFT and SURF are Scale invariant, which means that both
can detect features for images captured from different distances (affecting the size of
a target, in a picture). The HOG does not provide this important feature. The SIFT
algorithm provides another critical feature for our purposes: it is rotation invariant,
meaning that it can contour rotation problems. Due to this second invariance feature
it takes a little more time to process an image, compared to the SURF algorithm [ 5 ].
Our approach relies on a set of characteristic points stored in a reference data-
base. Each point is designated by Point of Interest (POI) and consists of one or more
geo-referenced model images of a building, road or monument. In our approach we
assume that a mobile system is near the location of a POI if the POI is recognized in
one of the captured images assigned to it. Hence, whenever GPS signal fails, the
CV system takes over the geolocation data feed process: summarizing, captured
frames of the mobile system surrounding landscape are analysed by a CV algo-
rithm, in real-time, trying to recognize POIs (from the database) in the captured
frames. To the extent of the test cases, the current work demonstrates that, without
GPS signal and with the help of simple computer vision algorithms, it is possible to
obtain conclusive answers about a mobile system current location based on the
proximity to a well-known (POI) location.
In Chap. 2 , the proposed approach is described and explained in more detail and,
in Chap. 3 , the prototype used for tests is presented. Chapter 4 shows the tests that
we conducted in order to validate our prototype and Chap. 5 presents the achieved
results. In Chap. 6 6 the results are discussed and

finally Chap. 7 and Chap. 8

present the conclusions and future work perspectives.

2 Improving Geolocation by Combining GPS
with Image Analysis

The following method is based on the described approach which combines computer
vision with GPS (CV-GPS): conceptually, if we consider that the GPS function is to
assign coordinates to a location then, by using an inverse logic, if we see a particular
geo-referenced POI, then we can assume that we are near to the GPS coordinates of
that POI. In the city sightseeing tour example previously mentioned, the

“ near ”

concept to a POI can be the line of sight proximity inside an urban canyon.

2.1 CV-GPS Method

GPS alone cannot provide an accurate positioning in situations where the receiver
fails to see the required satellites. Feature recognition through computer vision
algorithms cannot be used as a solo geolocation method because it would imply to
compare millions of images, trying to identify a feature (with geolocation previ-
ously associated). For current microcomputers, such task is not possible to be