Multi-target & multi-detector people tracker for mobile robots

摘要

People tracking is a key perception skill for mobile robots designed to share environments with human beings. It allows the robot to keep track of people around them, which is fundamental for two main reasons: safety and social interaction. This paper presents the work done on people tracking with the REEM robot after two years of paticipation at the RoboCup@home challenge. The main contribution of the paper is the tracker part, which is designed to be multi-target and to fuse heterogeneous detections from a variety of sensors, each one yielding different rates, field of views and quality performance. The paper carefully describes the tracker approach, based on multi-target particle filtering, as well as data association step, based on a probabilistic multi-hypothesis tree. Quantitative evaluations of real datasets using CLEAR MOT metrics are provided, comparing different sensor/detector set-ups and different data association approaches.