This paper reports on a novel technique to visually detect loop closings in feature-poor underwater environments in order to increase the accuracy of vision-based localization systems. The main problem of the classical visual Simultaneous Localization and Mapping (SLAM) for underwater vehicles is the lack of robust, stable and matchable features in certain aquatic environments. The presence of sandbanks, seagrass or other underwater phenomena cause the visual features to concentrate in regions heavily textured, leaving great image areas completely free of visual information. In this situation, the classical loop closing detection algorithms fail, resulting in no corrections for the SLAM system. Our novel method proposes to reinforce the loop closing detection by clustering visual keypoints present in multiple keyframes and to match features of clusters instead of features of keyframes.
This new technique is assessed on the particular application of navigating an Autonomous Underwater Vehicle (AUV) in marine environments colonized with seagrass or with the presence of sandbanks. Experiments conducted in several coastal zones on the Balearic Islands show a high degree of success in the visual registration of overlapping areas.