Real-time Hash-based Loop Closure Detection in Underwater Multi-Session Visual SLAM

This paper presents the assessment of a visual hash-based loop-closing detection method applied on marine environments colonized with seagrass. The loop-closing task is part of a complete Multi-session Simultaneous Localization and Mapping (SLAM) module executed online and on board an underwater autonomous vehicle. The loop-closing procedure is crucial for a correct and accurate map joining and has several strong points with respect to other visual loop-closing approaches: a) it has been applied in underwater areas extendedly colonized with seagrass, with highly successful results, despite the extremely complex textures of these kind of marine environments, b) it runs online, during the robot motion without hampering or delaying the entire navigation or localization process, and c) the use of hashes to characterize images and to find those that close loops reduces considerably the amount of inter-session data to be exchanged when the localization processes are run on board different vehicles. Extensive tests using visual datasets grabbed with an AUV in Mediterranean marine ecosystems colonized with seagrass show the suitability of the method in this type of environments.