Stereo SLAM for robust dense 3D reconstruction of underwater environments

Abstract

The virtual reconstruction of underwater environments in 3 dimensions can be of great utility for scientific or industrials applications, being the use of Autonomous Underwater Vehicles (AUVs) equipped with cameras an invaluable tool with progressive improvements. However, a highly accurate vehicle localization process is fundamental to place every portion of that 3D model in its real position with respect to the origin of the world system of coordinates, and to join properly the overall reconstructed area, especially in large surveys. Simultaneous Localization And Mapping (SLAM) techniques constitute the most precise localization approach using only data provided by the navigation sensors installed on board the underwater robot. However, the real challenge consists in applying these techniques in underwater environments where the imaging conditions are usually limited or degraded. This paper presents the comparison of two different SLAM approaches based on stereo-vision, a graph-SLAM and an EKF SLAM, applied to localize, in real time, an AUV moving in underwater environments. Both algorithms deal with pure 3D data, (x, y, z) for the vehicle position and a quaternion to represent its orientation. The aim of this work is to assess and compare the performance of both solutions in terms of accuracy in the estimation of the robot pose. First experiments were conducted in controlled aquatic scenarios where it was feasible to build a highly reliable ground truth, showing how the graph-SLAM approach outperforms its EKF counterpart under the same working and environmental conditions. A second set of experiments was conducted in the sea, showing the same tendency in the results. The SLAM pose estimates, given with respect to a global world frame, corresponding to the approach with less accumulated error are used to recreated virtual 3D maps of the environment based on the concatenation of successive stereo visual point clouds placed in the corresponding SLAM locations.