Multi-Camera Multi-Person Tracking Using Reinforcement Learning

Sammanfattning

The problem of multi-object-tracking in a network of cameras is an interesting and non-trivial problem. Given videos from a number of cameras the goal of Multi-Camera Multi-Object Tracking (MCMOT) is to find the full visible trajectory of each pedestrian from the videos as the pedestrians move across cameras. Compared to monocular tracking the main difficulties lie in associating pedestrian detections to tracks in locations where the camera views overlap. We develop two MCMOT methods; a traditional threshold-based Simple Online and Realtime Tracking (SORT) algorithm and a reinforcement learning track management method. The methods are evaluated on Blender 3D simulated data sets and on real-world recordings from calibrated cameras. The reinforcement learning model outperforms the more traditional method in Multi-Object Tracking Accuracy (MOTA) up to 4 % on the simulated and real data sets. Finally, our experiments show that the reinforcement learning method can transfer knowledge from simulated to real data during inference.