Assessing Image Segmentation Algorithms for Sky Identification in GNSS

In order to improve the accuracy of user's position solution using Global Navigation Satellite System (GNSS) in urban canyons, it is important to know whether a satellite's signal is obstructed by surrounding buildings. This can be accomplished by using an upward-facing camera and segmenting the image into sky and non-sky. This paper evaluates the Otsu, Mean Shift, Graph cut and HMRF-EM-image image segmentation algorithms for this purpose. Since some algorithms provide two or more categories, segmentation is followed by k-means clustering techniques to yield only two categories; sky and non-sky. The algorithms are tested using images taken using an upwardfacing camera at roughly the same locations in different weather conditions: cloudy and sunny. Result shows that, when images are appropriately adjusted, the Otsu method overcomes the three other algorithms in terms of the percentage of sky accurately segmented and is also more computationally efficient. Experiment was also perform in Calgary downtown to show the effect of segmentation on the GNSS accuracy. Results show that, when obstructed satellites are removed, the RMS of the residuals decreases significantly compare to when all satellites are used.