Anomaly detection in surveillance videos based on dual prediction networks

The field of computer vision has witnessed a surge of interest in the realm of surveillance video anomaly detection, a subject that has garnered significant attention from researchers and practitioners alike. This area of research boasts a diverse array of potential applications, spanning various domains and disciplines. Video anomaly detection based on single prediction networks is not sensitive enough to scale changes in the target and small-scale targets in the surveillance scene. The reconstruction error of the anomalous target is minimal, which results in leakage of the anomaly detection system and reduced performance in detecting anomalies. In order to address the shortcomings of small reconstruction error, target scale change and small-scale targets, this paper proposes a dual prediction network-based surveillance video anomaly detection method. The method increases the sensitivity to scale change and small-scale targets through the second prediction network, and increases the reconstruction error of anomalous targets. It is evident that the motion constraints, which are based on LiteFlowNet optical flow estimation and Gauss-Laplace loss function, are utilised during the training phase. This is done with the objective of enhancing the reconstruction performance of the predicted frames. The effectiveness of the proposed method and the improved detection performance of the video anomaly detection system are experimentally verified.