Background Suppressed and Motion Enhanced Network for Weakly Supervised Video Anomaly Detection

Yang Liu, Wanxiao Yang, Hangyou Yu, Lin Feng, Yuqiu Kong, Shenglan Liu

Abstract: Weakly supervised video anomaly detection (VAD) is a significant and challenging task in the surveillance video analysis field, locating anomalous motion frames using only video-level label information. In general, weakly supervised VAD mainly faces two obstacles. Firstly, the greatly changed background always causes false detection. Secondly, the motion changes are implicit in real-world surveillance videos caused by unsuitable focal distance, illumination, etc. This paper proposes a background suppressed and motion enhanced network (BSMEN) for Weakly Supervised VAD to solve the problems. The BSMEN utilizes the multi-head self-attention mechanism to construct a triple branch framework to suppress the influence from the background and enhance the motion significance in the VAD process. Moreover, a motion discrimination sequence extraction (MDSE) module is devised to locate the anomaly motion frames more accurately in the BSMEN. Extensive experiments on two mainstream VAD evaluation datasets validate that BSMEN outperforms state-of-the-art weakly supervised VAD methods.