Deep Reinforcement Learning-Based Multisensor Control for Labeled Multi-Bernoulli Filtering
Yue Yu, Mei Liu · IEEE Transactions on Aerospace and Electronic Systems · 2025
In this paper, we present a sensor control algorithm designed for directional multi-sensor systems. Each sensor, while fixed in position, possesses the ability to adjust its orientation and operates autonomously as an independent agent. By leveraging a deep reinforcement learning (DRL) approach, each sensor optimizes its actions to enhance the tracking of detected targets while simultaneously exploring for additional ones. Many existing sensor control methods rely heavily on predefined target motion models and struggle to effectively balance the tasks of searching and tracking, often falling short in achieving comprehensive real-time results. In contrast, DRL offers adaptive learning capabilities that are well-suited to address these challenges. Among DRL techniques, the multi-agent deep deterministic policy gradient (MADDPG) algorithm is particularly noteworthy for its ability to utilize global information during training to ensure stability, while making decentralized decisions based on local data, thereby meeting the requirements of distributed multi-sensor systems. This approach is integrated with the labeled multi-Bernoulli (LMB) filter for multi-target tracking, effectively managing target births, deaths, false alarms, and missed detections. Simulation results underscore the performance of the proposed algorithm, showcasing its effectiveness in demanding multi-target tracking scenarios.