A Deep Reinforcement Learning Visual Servoing Control Strategy for Target Tracking Using a Multirotor UAV

In this work, a deep Reinforcement Learning control scheme is developed in order to execute autonomous tracking of an unmanned ground vehicle (UGV) with a multirotor unmanned aerial vehicle (UAV). The UAV is equipped with a downward looking camera and the detection of the target UGV is achieved through a convolutional neural network (CNN). The deep RL control policy is deployed as a cascade position controller, which commands the inner attitude controller of the autopilot, and achieves the following of the UGV despite the aggressive maneuvers of the target. The efficacy of the proposed framework is demonstrated via a set of outdoor experiments using an octocopter flying above the ground vehicle.