Deep Reinforcement Learning Approach for Prioritized Heterogeneous Object Tracking in UAV Swarms

Deep Reinforcement learning(DRL) has recently showcased its boundless potential in Multi-agent collaboration. There exist DRL-based models that strive to unravel the Unmanned aerial vehicle (UAV) Multi-target tracking predicaments, however, the conundrum of tracing desired targets persists as most research endeavors solely focus on the agents’ policies while neglecting the dynamic environment, which harbors elusive and unpredictable targets. The proposed model executes Prioritized Heterogeneous Object Tracking (PHOT) using DRL, to furnish a solution for preferred Multi-target tracking by upholding the policy of UAV swarms. Tracking preferred targets assumes paramount importance as these targets necessitate meticulous monitoring within the ever-changing environmental backdrop, particularly during surveillance missions employing UAV swarms, which can harness the full potential of the UAV swarms in real-time. A YOLO v4-based algorithm is put forth, which can trail prioritized targets in a tracking by detection paradigm. To maintain the policy of UAV swarms, a dueling double deep Q-learning algorithm is deployed, complemented by a reward function that stimulates UAVs to track the preferred targets instantaneously. The UAVDT dataset is employed for the simulations, and the results evince that PHOT exhibits robustness, suitable for deployment in a real-time environment, amassing a highly commendable average cumulative reward over time, and enabling actions predicated on unidentical prioritized targets.