Kernelized Locality Sensitive Hashing for Fast Image Landmark Association

As the concept of war has evolved, navigation in urban environments where GPS may be degraded is increasingly becoming more important. Two existing solutions are vision-aided navigation and vision-based Simultaneous Localization and Mapping (SLAM). The problem, however, is that vision-based navigation techniques can require excessive amounts of memory and increased computational complexity resulting in a decrease in speed. This research focuses on techniques to improve such issues by speeding up and optimizing the data association process in vision-based SLAM. Specifically, this work studies the current methods that algorithms use to associate a current robot pose to that of one previously seen and introduce another method to the image mapping arena for comparison. The current method, kd-trees, is ecient in lower dimensions, but does not narrow the search space enough in higher dimensional datasets. In this research, Kernelized Locality-Sensitive Hashing (KLSH) is implemented to conduct the aforementioned pose associations. Results on KLSH shows that fewer image comparisons are required for location identification than that of other methods. This work can then be extended into a vision-SLAM implementation to subsequently produce a map.