Multi-Agent Reinforcement Learning as Interaction Model for Online Multi-Object Tracking

This paper presents fast and efficient method for Multi-Object Tracking. It uses Multi-Agent Reinforcement Learning techniques to create an Interaction Model capturing the dynamics of tracked objects that are treated as individual cooperative agents. Built on base of YOLO v5 (You Only Look Once) object detector and using state-of-the-art MADDPG (Multi-Agent Deep Deterministic Policy Gradient) reinforcement learning algorithm, the proposed method can achieve accurate real-time performance after offline training. Experimental results on challenging MOT16 and SeaDronesSee datasets are presented to demonstrate the effectiveness of the proposed method.