Collision Avoidance with Adaptive Potential Fields for Underwater Vehicles Using Omnidirectional Vision

Exploration of marine environments is commonly performed at close-range with optical sensors to acquire images of the underwater structures. However, underwater obstacle avoidance is usually performed using acoustic sensors that, when working in closer proximity of the obstacles, become less reliable. Operating an AUV or an ROV close to a structure requires a higher level of alertness to avoid possible collisions. In this paper, we present a collision avoidance algorithm that leverages the information from an omnidirectional multi-camera system to correct the robot's current path to maintain a safety distance from surrounding structures and that performs preventive manoeuvres to prevent future collision by the considering its current speed and velocity. The tracked image points from the cameras are triangulated and then optimised via bundle adjustment to generate a more accurate point cloud for the collision avoidance. The collision avoidance step constructs a repulsive potential field using the refined point-cloud, which is then used to veer the AUV towards the safest direction. The tests were performed in a realistic physical simulator for underwater robotics (on Stonefish [1]) as a first step towards validating the system at sea. Results of the testing show that the proposed approach is able to adequately steer the AUV away from potential collisions, and that is suitable to be used in realtime under the limited computing power available in autonomous underwater platforms. Future work includes extending it to a local planner that actively attempts to modify the robot's current trajectory on a fixed horizon window, and adaptation of the algorithm for real-world use.