This work presents a real-time multi-person or multi-object tracking algorithm that uses multiple hypothesis reasoning in time to enforce multi-person match constraints. The algorithm is intended to augment, not replace, existing multiperson tracking methods. We demonstrate how tracking systems that use inter-frame feature matching can be improved by enforcing contextual matching constraints throughout a 1-5 second temporal window. Robust and efficient multiple hypothesis reasoning in time is achieved for a useful class of tracking problems using a dynamic programming framework. Results are described for a dataset of 40 minutes of test video taken from a static, top-down camera and with two to four people moving about a small room.
Multi-person visual tracking, computer vision, temporal, dynamic programming.