Multi-memory video anomaly detection based on scene object distribution

With the popularity of surveillance equipment and the rise of intelligent surveillance, video anomaly detection has gradually become a research hotspot. Among them, for video processing, the three-channel video frame data can be directly used as the input of model, or some motion information can be extracted from the video frame, such as calculating optical flow, and then motion information and video frame can be input into the model together for anomaly detection. However, since the amount of background information in the overall situation is far greater than that of object information, abnormal objects are not concerned. In addition, there ia a phenomenon that objects close to the camera are more likely to be judged as anomalous due to the difference in viewpoint resulting in different sizes of objects captured in the scene. This paper proposes a multi-memory video anomaly detection algorithm based on scene object distribution. Firstly, add local anomaly branch to the model, and use memory modules to explicitly model the multiple normal modes of the global frame and the local object; secondly, scale the object to the same measurement standard according to the scene object distribution, which alleviates the impact of the view difference; finally, considering the difficulty of anomaly positioning, a new anomaly location method that combines global anomalies and local anomalies is proposed. The experimental results on the UCSD Ped2, CUHK Avenue and ShanghaiTech datasets have obtained AUC values of 96.75%, 84.34% and 77.08% respectively, which shows that the proposed method attains competitive detection accuracy.