Crowd Anomaly Event Detection in Surveillance Video Based on the Evolution of the Spatial Position Relationship Feature

While cameras are widely used in surveillance for public security, the abnormal crowd behavior detection becomes an important research issue in computer vision. The traditional behavior methods mainly focus on the detection of individual behaviors and rely on a fixed detection threshold that is learned from the sample set which leads to greater computational complexity. The performance is susceptible by the quantity and quality of samples and thus limits their applications. In this paper, a method based on the evolution of spatial position relationship (ESPR) features is proposed to detect there types of abnormal crowd behaviors in videos: crowd spread behavior, crowd gather behavior and crowd collective movement behavior. First, the optical flow features of each frame are extracted though an improved optical flow method. And then, the features are analyzed through fuzzy clustering to acquire the tightness level of all moving object. Finally, the abnormal event is detected based on the ESPR feature converted by tightness level frame by frame. In the proposed method, the learning stage is unnecessary. Experiments are conducted on public available datasets, namely, UMN dataset and Web dataset. The performance of the proposed method is evaluated in various scenarios.