Rapid detection of crowd abnormal behavior based on the hierarchical thinking

Abstract

In view of current algorithms for the crowd abnormal behavior detection are not suitable for different scenarios, a hierarchical detection algorithm is proposed. First, the target area is extracted using the Gaussian mixture model. According to the pixel ratio, decide which type the group belongs to: individualism behaviors, social interaction behaviors or leadership-led behaviors. Then divide the video according to the category of the crowd, calculate HOG-LBP for the crowd with individualism to judge the abnormal appearance. For other categories, calculate the trajectory and entropy in the divided image to obtain the speed, deviation from the trajectory and variance of the trajectory. Then compare with the corresponding thresholds to determine whether an abnormality occurs. The value of the entropy and its first-order function are used to judge the abnormal extent. When the entropy does not exceed 3/2 of the threshold, the optical flow is extracted to calculate CMI, and the peak value is used to detect anomalies. After experiments, our algorithm is verified to be rapid and accurate in different scenarios.