Video Anomaly Detection Based on Multiple Instance 3D Channel Attention

Abstract

Video anomaly detection is a hot topic in the field of computer vision and has broad application prospects. To address the issues of feature enhancement and temporal continuity, this paper proposes a framework named multiple instance 3D channel attention. In particular, this framework includes two networks, i.e., the pseudo label generation network and the anomaly detection network. Based on smoothness and sparsity constraints, pseudo labels for video clips are generated in the pseudo label generation network. Moreover, the 3D channel attention block is designed to enhance features. On the ShanghaiTech dataset, experimental results demonstrated that the proposed method obtained better performance than baseline method and other methods.