Video Anomaly Detection Based on Convolutional Neural Network
Asia-pacific Journal of Convergent Research Interchange · 2022
The intelligent video surveillance technology, which can analyze and report the abnormal events that endanger the social order real time, has been widely used.As an important research direction of intelligent video surveillance technology, video anomaly recognition focuses on how to extract video features and process them and detect anomalies.The features used in traditional video anomaly detection algorithms are mainly motion features based on manual design.However, the characteristics of artificial design need some prior knowledge, which mainly depends on the monitoring target and is difficult to define in different applications.Based on the feature that convolutional neural network can automatically learn the features related to tasks during training, this paper proposes a video anomaly detection algorithm.In video prediction network, a self-coding structure is proposed to realize video feature extraction, timing modeling, and prediction frame generation.The 3DCNN and LSTM are combined to enhance the network's ability to model video actions, and the feature images are fused by stitching to reduce the information loss in the process of feature extraction.After that, a multi-task video prediction model is designed based on the action features extracted from the non-deep learning model.With the help of MOG2 to extract the foreground image and Farneback to extract the optical flow image, the network can predict the video frame and the action features of the video frame at the same time.In addition, according to the action features extracted from the deep learning model, another multi-task video prediction model is designed.With the help of Flownet2-SD, the optical flow graph is extracted , and the video prediction network is optimized by minimizing this difference.The effects of different auxiliary tasks on video prediction ability and video anomaly detection ability of different data sets are discussed.