LAUSR

Automatically and efficiently detecting abnormal events that occur in dynamic surveillance video is one of the important tasks in real-time video streaming analysis. However, in some typical application scenarios with narrow areas, the perspective distortion caused by the large depth-of-field has a tremendous negative impact on the accuracy of detection, thereby increasing the difficulty of identifying abnormal behavior. Taking the real-time violence occurring in the metro platform as an example, the article introduces a more effective algorithm for detecting abnormal behaviors in narrow areas with perspective distortion. The algorithm firstly uses the adaptive transformation mechanism to make up for the distorting effect in the region of interest extraction. Then, an improved pyramid L–K optical flow method with perspective weight and disorder coefficient is proposed to extract the abnormal behavior feature occurred in historical moving images. The side-by-side comparison of the experimental results proves that this algorithm can effectively compensate for the distortion effect and obviously improve the accuracy of abnormal behavior detection in narrow area scenes.