Abnormal behavior detection based on spatio-temporal information fusion for high density crowd

Abstract

With the strong demand of real-time detection of crowd abnormal behavior, an algorithm of abnormal behavior recognition based on video sequence for high-density crowd without any training stage is proposed, aiming at realizing online real-time detection of crowded events in intelligence video surveillance system. The problem of blurred targets and limited pixels of target expression will make it difficult to segment targets, which will affect the recognition of group behavior. In addition, severe occlusion limits the performance of traditional visual tracking methods. The use of classifier will result the dependence of scenes, and even affects the online deployment of intelligence video surveillance system. In the stage of detection process, the new method includes motion region segmentation, motion blobs extraction and crowd activity image creation, and then the information entropy of crowd activity image is used for the recognition of crowd behavior. The new method carries out experimental and quantitative data analysis on selected public data sets, and com-pares it with other online crowd event detection methods. The experimental results show that the method can detect ab-normal crowd behavior in high-density situations without training stage, and has better detection effect than the state-of-the-art technology.