Robust Active Visual Tracking of Space Non-Cooperative Objects

Deep Reinforcement Learning (DRL) based Active Visual Tracking (AVT) algorithms targeting Space Non-cooperative Objects (SNCOs) is very vulnerable to various perturbations such as temporal action control command failure, actuator failure or signal transmission failure. Such perturbations can severely affect the performance of active visual trackers. Thus in this paper, targeting action failure, a robust DDPG based AVT algorithm is proposed which uses a new reward function to prevent DRL over-fitting. The proposed algorithm shows resistance to the perturbation and is able to perform outstanding tracking under high action failure probability. Sufficient experiments were conducted to verify the effectiveness and advancement of the proposed algorithm.