DOSeqSLAM: Dynamic On-line Sequence Based Loop Closure Detection Algorithm for SLAM

Abstract

Simultaneous Localization and Mapping (SLAM) is vital for modern autonomous robots. Visual place recognition of pre-visited areas, widely known as Loop Closure Detection (LCD), constitutes one of the most important SLAM components. In this paper a sequence-based LCD algorithm is proposed by evolving SeqSLAM method. Instead of using fixed-size sequences' length during search process, as in the original approach, we suggest a dynamical length definition based on the images' content proximity. Specifically, sequence segmentation is achieved through a feature matching technique applied on the incoming visual sensory information, while the mechanism operates on-line without the need of any pre-training procedure. Each sequence's similarity score, produced by a weighted average function, is utilized as a decision factor for the loop closing selection. Finally, an extended image-to-image search in the chosen group of images, avails the system of identifying the appropriate match. A temporal constraint prevents early pre-visited areas to be presented as false-positive matches. The method is evaluated on several outdoor publicly-available datasets, revealing a substantial improvement on SeqSLAM.