Video anomaly detection algorithm based on knowledge distillation and deep probability estimation

Video anomaly detection aims to automatically identify anomalous targets, behaviors, or events in surveillance video clips. Most of the current work focuses solely on learning features of normal samples from a limited number of training sets, which hinders the ability to glean prior knowledge from large-scale datasets. Moreover, anomalies are typically defined only at the pixel level, which may result in missed detection. Therefore, we propose a framework for video anomaly detection based on knowledge distillation and deep probability estimation. In this paper, we leverage knowledge distillation to incorporate prior knowledge from pre-trained models in the field of video anomaly detection. This approach enhances the differentiation between normal and anomalous samples and increases the separation in the feature space based on prior knowledge. Additionally, we introduce a deep probability decision model that constructs hyperspheres in the feature space to delineate anomalies. We evaluated the proposed anomaly detection model on several popular video surveillance datasets (such as UCSD, Avenue, ShanghaiTech, etc.) and achieved competitive performance compared to state-of-the-art methods.