Human Motion Analysis Using Simultaneous Trajectory and Body Detection and Modeling

In this paper, we propose a scheme to combine trajectory-based detection and body-based estimation to analyze human behavior in video scene. Our scheme is applied for a fast and automatic detection of pick-up/drop-off events within surveillance videos in indoor areas. A moving person is tracked globally using the mean-shift algorithm and modeled locally using an axis skeleton in a monocular video sequence. We detect and rectify the stationary pick-up/drop-off objects. The spatial-temporal relationship between object and person is measured and exploited to detect a pick-up/drop-off event. Our experimental results accurately estimate the human-motion trajectory and infer the posture. The system operates at real-time speed (around 20 frames/second).