Abnormal Crowd Behavior Detection Using Speed and Direction Models

This paper presents a novel method to detect unusual crowd behavior in a video sequence using probability models of speeds and directions. Thus, optical flow is used to extract velocities at each image frame, which are then reduced to speed and motion orientations. Using expectation maximization algorithm, we construct a mixture model of von Mises distribution describing the set of directions, and a mixture model of normal distribution related to the speed set. Each frame is compared with a collection of reference frames using distance of probability densities. This distance is then used to indicate changes in the crowd motion. Unlike the speed based detection, using the direction model is not yet adapted to the case of unstructured crowds. The proposed method was tested on various publicly available crowd datasets.