Online Pedestrian Tracking Using A Dense Fisheye Camera Network With Edge Computing

This paper describes a dense fisheye camera network for pedestrian tracking in an indoor environment, with the target scenario being an unmanned store. Each fisheye camera is connected to a single-board computer for local tracking using its own images. The local tracks are integrated and global tracks generated at a central computer in an online manner. The local trackers are based on the popular DeepSORT algorithm, and the global tracker combines distance and novel specialization based factors to update global tracks from local tracks, avoiding the need of matching local tracks. Experiments on a self-collected dataset demonstrate highly accurate tracking over several minutes of videos.