A Distillation Network using Improved ConvLSTM for Video Anomaly Detection
Jialong Li, Chunjie Zhou, Shengjie Liu, Xiaoyun Lu · 2024
Video anomaly detection task has significant research and application value in areas such as intelligent security, traffic monitoring, and industrial manufacturing. Due to the scarcity and unlabeling of abnormal samples, existing methods usually rely on the features of data itself. However, for abnormal samples with slight differences from normal samples, the strong generalization of most existing networks based on reconstruction or prediction cannot guarantee a obvious distinction for such anomalies. Recent studies have proved that knowledge distillation is an effective approach to solve this problem. This paper introduces knowledge distillation technology within the video prediction network, which avoids the problem of false detection of anomalous data due to lack of clear features. Firstly, the teacher network is pretrained on a large action video dataset. And the student network introduces Mogrifier gating in traditional ConvLSTM to enhance the modeling of contextual information. Due to the different observation fields of student and teacher, student can get significantly different output from teacher when faced with abnormal data. To strengthen the consistency between teacher-student’s output, this paper employ intensity and gradient constraints. The proposed method is validated on the ShanghaiTech and UCSD Ped2 datasets, and experiments show that the proposed approach exhibits advanced performance in robustness to normal events and sensitivity to anomalous events.