A novel approach for global abnormal event detection in multi-camera surveillance system

In this paper, a novel global abnormal event detection algorithm is proposed for multiple disjoint synchronous camera network. We treat detecting unusual global events as discovering context-incoherent patterns through learning temporal dependencies between distributed local activities observed within and across camera views. Trajectories are firstly extracted using mean-shift approach in each camera view. As local activities are learned by applying clustering algorithm for trajectories, we model global event patterns using a probabilistic graphical model with different nodes representing entry/exit regions from different views and the directed links between nodes encoding their temporal dependencies. A novel two-stage structure learning algorithm is formulated to learn globally optimized temporal dependencies. Modified Dynamic Time Warping is used to learning the links in unobservable regions in camera network. Then, Monte Carlo(MC) algorithm is used to refine the structure and produce a final dependency structure. We validate the effectiveness of the proposed approach using a synthetic data set and videos captured from a camera network installed at a research institute.