Development of a navigation system using Visual SLAM

Abstract

Supporting the work of the Dutch company Almende B.V., this work has developed an autonomous navigation planner for an underwater unmanned vehicle. This vehicle is tasked with performing wall inspection using a video camera. Usually, this inspection job is performed by a remote human operator, this work has looked at reducing the efforts of the operator. A navigation challenge like this requires a high-level control loop, consisting of a video camera, a visual SLAM algorithm, a global navigation planner, a control system, and actuators on the vehicle. For the purpose of this work, many of these components were taken off-the-shelf, however, the SLAM algorithm and navigation planner needed developing. The strategy for navigation was based on the quality of the already acquired inspection images, with the vehicle actively trying to move towards areas with low-quantity and low-quality images. An important step to achieve this was the introduction of a quality map, an Octomap-based representation of this quantity and quality metric. The actual implementation was tested through simulation within Gazebo (using a marine environment plugin). It has been shown in the simulated environment, that this strategy is feasible, however, several enhancements can still be applied. Almende B.V. is planning on applying the results of this work in a future proof of concept with a real underwater inspection robot, in the Rotterdam Harbor area.