Abnormal Event Detection Method in Surveillance Video Based on Temporal CNN and Sparse Optical Flow
Hongxia Xia, Ting Li, Wenxuan Liu, Xian Zhong, Jingling Yuan · 2019
Abnormal event detection in surveillance video is very important for intelligent analysis of video. Most anomalous event detection methods rely on complex feature extraction to represent motion and appearance. Convolutional Neural Network (CNN) is a powerful and efficient tool that can fully meet the needs of feature extraction. In this paper, tracking CNN features over time can effectively detect local anomalies. We detect local anomalies by combining temporal CNN models with optical flow. Based on the traditional optical flow method, the foreground mask is used to improve the efficiency of optical flow calculation and the robustness of the optical flow. The method was tested and evaluated on real-life surveillance video and benchmark datasets to verify its effectiveness.