Reinforcement Learning for UAV Surveillance Systems

Abstract

In this paper, the authors focus on a surveillance system with unmanned aerial vehicles, UAVs. Moving objects are targets in this system. In order for UAVs to detect the moving objects, the UAVs are required to take the optimal action depending on the situation. For this purpose, we adopt the reinforcement learning approach. The situation of the UAVs and targets are defined as a state. Given a target that moves sufficiently faster than the UAVs, however, it is difficult for the UAVs to learn the optimal action. For this challenge, we design a novel reinforcement function based on the expected value of reward obtained in each state. In order to prevent the state space from increasing, we further define the state of the UAVs and target in the relative position coordinate system. Surveillance simulations show that the UAVs are enabled to detect more targets.