Tracking of Multiple Humans in Meetings

Tracking conferees in meeting rooms is important for many applications. In this paper, we present an approach based on single-frame head-shoulder detection to track multiple humans in meetings. The responses of a multiple view head-shoulder detection system are taken as the observation of the human hypotheses. Trajectory initialization and termination are fully automatic and rely on the evidence collected from the detection responses. An object is tracked by data association if its corresponding detection response can be found; otherwise it is tracked by a meanshift style tracker. Finally the tracked hypotheses are verified by evidence collected from body part movements. The system is evaluated on two meeting video corpora.