Towards Multi Session Visual SLAM in Underwater Environments Colonized with Posidonia Oceanica

Abstract

[eng] This paper describes a multi-session monocular SLAM approach addressed to underwater environments. It has three main blocks: a) visual odometry, b) loop-closing detection; loop closings inside each individual session are found applying feature matching with RANSAC, and since no geometric relation between images of different sessions is available, multi-session loop closings are detected via an image hash matching procedure, alleviating also the computational cost of the image comparisons, and c) an Iterated Extended Kalman Filter (IEKF) based optimization process used to refine the different individual trajectories and to join the different maps; the global optimization process (map joining) can be delayed until certain number of loop closings are found, reducing the global running time. Several experiments using marine imagery in areas colonized with Posidonia Oceanica show the robustness of this approach.