Image segmentation for automated taxiing of Unmanned Aircraft

This paper details a method of detecting collision risks for Unmanned Aircraft during taxiing. Using images captured from an on-board camera, semantic segmentation can be used to identify surface types and detect potential collisions. A review of classifier lead segmentation concludes that texture feature descriptors lack the pixel level accuracy required for collision avoidance. Instead, segmentation prior to classification is suggested as a better method for accurate region border extraction. This is achieved through an initial over-segmentation using the established SLIC superpixel technique with further untrained clustering using DBSCAN algorithm. Known classes are used to train a classifier through construction of a texton dictionary and models of texton content typical to each class. The paper demonstrates the application of said system to real world images, and shows good automated segment identification. Remaining issues are identified and contextual information is suggested as a method of resolving them going forward.