Collaborative multi-target-tracking via graph-based deep reinforcement learning in UAV swarm networks

Abstract

Due to unmanned aerial vehicles (UAVs) flexibility and affordability, the UAVs swarm network (USNET) is widely used for various complex, challenging tasks such as tracking, surveillance, and monitoring, and the key to accomplishing these tasks lies in the capabilities of the UAVs to collaborate. However, due to the high complexity of real-time information sharing and task cooperation among numerous UAVs in the USNET, it poses significant challenges for multi-target tracking in complex scenarios. In this paper, we study the collaborative multi-target-tracking (CMTT) problem based on the USNET and aim to improve task collaboration capabilities within the USNET. We first design a heuristic target assignment algorithm to simplify the CMTT problem into the optimal topology control problem of the USNET, and then propose an integrated sensing and communication multi-agent reinforcement learning for the USNET topology control algorithm (ISAC-TC) to maximize the collaborative tracking performance of UAVs within the USNET. Specifically, in heterogeneous observation graph representation, the ISAC-TC first utilizes a graph neural network to solve the time-varying dimensions of the agent observation space. Then, an encoder–decoder-based information sharing module is used to achieve efficient communication between agents in the CMTT tasks. Simulation results show that the proposed scheme achieves a higher tracking success rate and tracking fairness than other baselines.