Video-Based Detection of Abnormal Behavior in the Examination Room

Aiming at the problems of partial occlusion and background clutter in the examination room, we propose a method for behavior detection using spatial-temporal shape and flow correlation. The method first extracted training templates using interactive video cutout technique, and automatically segmented the video into 3D spatial-temporal volumes using improved Mean Shift algorithm. Then we slide the template across the video and compute the matching distance. We complement our shape-based features with flow, and efficiently match the volumetric representation of an action against over-segmented spatial-temporal video volumes. Thresholding the correlation distance and finding the peaks give us locations of potential matches. The experiment results indicate that this method achieves human's action detection robustly in crowded, dynamic environment.