UMD_VDT, an Integration of Detection and Tracking Methods for Multiple Human Tracking

This paper presents a cooperative body part labeling and tracking algorithm for locating body parts in a videoimage sequence. The body part labeling procedure is used to initialize the tracking procedure and to prevent
tracking from getting lost; and the tracking procedure is used to analyze frames that cannot be handled reliably by
the labeling procedure. This results in an automatic and robust body part tracking system. Unlike previous work
that always start tracking from the first frame which may be difficult to analyze, our tracking procedure starts
with some key frames in which body parts can be located reliably. Key frames are selected based on a labeling
confidence measure which evaluates the similarity between the human model and the image features. The labeled
body parts are tracked in-between frames based on such cues as shape, color, and location interpolated between
key frames iteratively. The proposed algorithm has been applied to analyze NASA videotapes in which crew
members present challenging postures of bending, twisting, and rotating in a space station. The experimental
results demonstrate the effectiveness of the proposed algorithm.