Feature Enhancement Based Multi-Feature Fusion Network for Video Anomaly Detection in Offshore Surbeillance

Jiacheng Gao, Xu Cheng, Xiufeng Liu, Fan Shi, Meng Zhao, Shengyong Chen

Abstract: With the advancement of deep learning techniques and the widespread use of surveillance systems, there is an increasing demand for detecting anomalous events in offshore video scenes. This paper proposes a novel Multi-Feature fusion network (MFFN) based on video features. Firstly, we enhance the information entropy feature extracted from the video to generate the Video-Momentum-Feature (VMF). The VMF enlarges the difference between the entropy feature of normal and anomalous videos, and allows the network to focus on the most anomalous parts of the video. Moreover, the network employs I3D to extract the RGB and Optical flow characteristics separately. Then, the VMF is fused with these RGB and Optical flow characteristics, respectively. Finally, the fused features are utilized for anomaly detection. Experimental results on the modified UCF -Crime dataset and the offshore ferryboat dataset demonstrate that our proposed method achieves significant performance.