Active Visual SLAM with Exploration for Autonomous Underwater Navigation.

[摘要] One of the major challenges in the field of underwater robotics is the opacity of thewater medium to radio frequency transmission modes, which precludes the use of a globalpositioning system (GPS) and high speed radio communication in underwater navigationand mapping applications. One approach to underwater robotics that overcomes this limitation is vision-based simultaneous localization and mapping (SLAM), a framework thatenables a robot to localize itself, while simultaneously building a map of an unknown environment. The SLAM algorithm provides a probabilistic map that contains the estimatedstate of the system, including a map of the environment and the pose of the robot.Because the quality of vision-based navigation varies spatially within the environment,the performance of visual SLAM strongly depends on the path and motion that the robotfollows. While traditionally treated as two separate problems, SLAM and path planningare indeed interrelated: the performance of SLAM depends significantly on the environment and motion; however, control of the robot motion fully depends on the informationfrom SLAM. Therefore, an integrated SLAM control scheme is needed—one that can direct motion for better localization and mapping, and thereby provide more accurate stateinformation back to the controller.This thesis develops perception-driven control, an integrated SLAM and path planningframework that improves the performance of visual SLAM in an informative and efficientway by jointly considering the reward predicted by a candidate camera measurement, alongwith its likelihood of success based upon visual saliency. The proposed control architecture identifies highly informative candidate locations for SLAM loop-closure that are alsovisually distinctive, such that a camera-derived pose-constraint is probable. Results areshown for autonomous underwater hull inspection experiments using the Bluefin RoboticsHovering Autonomous Underwater Vehicle (HAUV).

[发布机构] University of Michigan

[效力级别] Mechanical Engineering

[关键词] SLAM AUV Path Planning Computer Vision Navigation;Mechanical Engineering;Engineering;Mechanical Engineering