Loop Closure Detection with Incremental Proximity Graph and Multi-words Quantization

This paper studies the visual loop closure detection problem. The online loop closure detection algorithm is proposed embedded with an incremental visual words codebook. The main novelty of the work lies: 1) To mitigate the coherent quantization error of feature words, a multi-words quantization approach is designed. 2) A hierarchical navigable small world graph is utilized as the codebook structure to alleviate the deficiency of high dimensional search in space partition methods. 3) A complete loop closure detection framework based on tracked word elements is tested and analyzed on several public image datasets. The results validate the computational efficiency of this algorithm.