Detection of Anomalous Crowd Behavior Based on the Acceleration Feature

In this paper, we propose a novel algorithm based on the acceleration feature to detect anomalous crowd behaviors in video surveillance systems. Different from the previous work that uses independent local feature, the algorithm explores the global moving relation between the current behavior state and the previous behavior state. Due to the unstable optical flow resulting in the unstable speed, a new global acceleration feature is proposed, based on the gray-scale invariance of three adjacent frames. It can ensure the pixels matching and reflect the change of speed accurately. Furthermore, a detection algorithm is designed by acceleration computation with a foreground extraction step. The proposed algorithm is independent of the human detection and segmentation, so it is robust. For anomaly detection, this paper formulates the abnormal event detection as a two-classified problem, which is more robust than the statistic model-based methods, and this two-classified detection algorithm, which is based on the threshold analysis, detects anomalous crowd behaviors in the current frame. Finally, apply the method to detect abnormal behaviors on several benchmark data sets, and show promising results.