A Trajectory-Based Approach to Multi-Session Underwater Visual SLAM Using Global Image Signatures

Abstract

This paper presents a multi-session monocular Simultaneous Localization and Mapping (SLAM) approach focused on underwater environments. The system is composed of three main blocks: a visual odometer, a loop detector, and an optimizer. Single session loop closings are found by means of feature matching and Random Sample Consensus (RANSAC) within a search region. Multi-session loop closings are found by comparing hash-based global image signatures. The optimizer refines the trajectories and joins the different maps. Map joining preserves the trajectory structure by adding a single link between the joined sessions, making it possible to aggregate or disaggregate sessions whenever is necessary. All the optimization processes can be delayed until a certain number of loops has been found in order to reduce the computational cost. Experiments conducted in real subsea scenarios show the quality and robustness of this proposal.

Towards Multi-Robot Visual Graph-SLAM for Autonomous Marine Vehicles

“…(2) The approach starts by estimating the trajectory of each robot motion, separately, applying the trajectory-based visual-SLAM strategy included in the multi-session scheme of Burguera and Bonin-Font [46]. Let us refer to this step as the intra-session SLAM.…”

“…In our particular case, this displacement corresponds to the visual odometry calculated between consecutive images, and the images candidate to close a loop with the current image are found comparing the corresponding image hashes and confirmed by a RANSAC-based algorithm applied on a brute-force visual feature-matching process. The main difference with respect to [46] is that while in a multi-session localization procedure the trajectory of the currently running robot is joined to another trajectory already completed and available in its totality, now, in a Multi-robot scheme, both robots are moving at the same time to complete a mission in which both participate simultaneously. When joined, both trajectories are incomplete, and continue running until all robot missions are finished.…”

“…Every robot geo-localizes itself with respect to the origin of its own trajectory, but it has no knowledge about the origin of the other trajectories. By means of finding inter-session loop closings, i.e., images that show, partially or totally, the same area, but taken by two different robots in two different sessions running simultaneously, the maps of the two robots can be joined in a single one [46]. Due to this lack of geometrical relation between the two trajectories, the search of the loop closing candidates of one robot to close a loop with the last image captured by the other robot cannot be restricted to a certain area.…”

“…One way to alleviate this problem is reducing all images to global descriptors. As in [46], all images of both sessions are reduced to their HALOC global descriptor. The size of HALOC is fixed in 384 floats since the size of the used projective orthogonal vectorial space is 3 [22].…”

“…In our case, robustness has been qualitatively validated by two means, (a) comparing the resulted global maps with the mosaics obtained with BIMOS [54] and, (b) comparing the direct transforms between images that close loops, obtained with Algorithm 1 with the transforms between the graph-nodes related to the same loop closing images. These two points have been already used and validated in previous pieces of work [46].…”