Unmanned Aerial Vehicle Control Through Deep Reinforcement Learning and Computer Vision for Obstacle Avoidance

This research presents an approach for enabling an Unmanned Aerial Vehicle (UAV) quadrotor to autonomously avoid obstacle collisions in an unknown environment. The task taken hereby brings in additional challenges when compared to ground robots as it involves six degrees of movement with latitude change as one of the key aspects. The inclusion of latitude change gives more freedom to maneuver UAV than ground robots but at the same time involves more types of obstacles to avoid. Thus, techniques used in ground robots are inadequate to be used in this work. In addition to the aforementioned challenges the focus of the work is to achieve autonomous obstacle avoidance (OA) by the use of computer vision (CV) using the monocular camera mounted on the UAV. Current control methods that use monocular camera for UAV obstacle avoidance do not use latitude change to navigate through the environment. The proposed deep reinforcement learning based UAV obstacle avoidance approach uses computer vision to do the same. The key idea in the method is the use of depth maps of the environment and inclusion of latitude change of the UAV to avoid obstacles. The technique used here provides a method for power efficient and stable navigation of the UAV in unforeseen environments.