Underwater visual SLAM with loop-closure using image-to-image link recovery

This study develops an underwater visual simultaneous localization and mapping (SLAM) algorithm using a monocular vision as a major measurement sensor, focusing in particular on the loop-closure problem. Although most vision-based loop-closure approaches have been implemented by feature-based pairwise image matching, matching underwater images is generally a challenging task due to the limited number of feature correspondences or relatively small overlapping regions between the compared images. The lack of loop-closing events caused by these challenges can degrade the navigation and mapping performance of visual SLAM. For robust visual SLAM, a loop-closure algorithm that can recover image-to-image matching links is presented. The proposed loop-closure algorithm is more resilient to failures of pairwise matching and thus can maximize the use of image-to-image links, thereby improving the estimation performance in the context of visual SLAM. To validate the effectiveness of the proposed algorithm, a hover-capable unmanned underwater vehicle was used for in-water experiments, and the proposed algorithm was also evaluated through a series of comparative results drawn from an experimental dataset.