Application of Locality Sensitive Hashing to Feature Matching and Loop Closure Detection

My thesis focuses on automatic parameter selection for euclidean distance version of Locality Sensitive Hashing (LSH) and solving visual loop closure detection by using LSH. LSH is a class of functions for probabilistic nearest neighbor search. Although some work has been done for parameter selection of LSH, having three parameters and lack of guarantees on the running time, restricts the usage of LSH. We propose a method for finding optimal LSH parameters when data distribution meets certain properties.;Loop closure detection is the problem of deciding whether a robot has visited its current location before. This problem arises in both metric and visual SLAM (Simultaneous Localization and Mapping) applications and it is crucial for creating consistent maps. In our approach, we use hashing to efficiently find similar visual features. This enables us to detect loop closures in real-time without the need to pre-process the data as is the case with the Bag-of-Words (BOW) approach.;We evaluate our parameter selection and loop closure detection methods by running experiments on real world and synthetic data. To show the effectiveness of our loop closure detection approach, we compare the running time and precision-recalls for our method and the BOW approach coupled with direct feature matching. Our approach has higher recall for the same precision in both sets of our experiments. The running time of our LSH system is comparable to the time that is required for extracting SIFT (Scale Invariant Feature Transform) features and is suitable for real-time applications.