A graph analysis method for abnormal crowd state detection

An abnormity detection method for a dynamic crowd scene is proposed based on graph analysis. After the non-parametric clustering in velocity space via an adaptive mean shift algorithm, we get the clustering results containing some cluster centers and Euclidean distances between them, and they can form a graph whose vertexes are the cluster centers and edge weights are the distances. Through analyzing the vertexes' distribution in feature space and the state transform of a dynamic system made by the sequence of the edge weight matrix, we can detect and locate the abnormal events in the scenario. To testify the method's effectiveness, we conducted experiments on several well-known datasets and obtained good performance in both abnormal events detection and location. The results show that the graph analysis method has strong adaptability and can efficiently detect the abnormal states in dynamic crowd scene.