Closed-loop person tracking and detection

We present a system that detects people in indoor scenes by modeling the motion history of foreground blobs, rather than their shape or appearance. The system tracks all foreground blobs over time with a multi-hypothesis tracker, and considers a blob to be a person if it exhibited sufficient autonomous movement in the course of its tracking history. This way, people can be correctly classified even if they are seen in a wide range of body poses, if they remain still for a long time, or if they change appearance by taking off a coat. Evaluation on over 1h of video demonstrated good performance for both heuristic and decision tree based classification.