Research And Implementation Of Video Anomaly Detection Based On Generative Model

With the increase of the density of video surveillance system,the task of video anomaly detection relying on human resources becomes more and more difficult.It has become an urgent need to achieve a higher level of intelligence of video surveillance system and enable the system to detect anomalies automatically.For the video anomaly detection task,in view of the difficulty in obtaining abnormal data,the discriminant model is usually not used,but the generated model is used to model the sample distribution of normal training datasets,and then the anomaly in video is detected according to the difference between test samples and training patterns.However,when the existing methods are used to model normal patterns,there is an imbalance between time information and spatial information,which leads to a preference for specific types of anomaly detection.This paper designs and implements video anomaly detection algorithm based on generative model.The main results of the paper are as follows:(1)We experiment the principle of using optical flow as a model input,and designed a data preprocessing method for model input.(2)We propose a neural network framework to better combine spatial and temporal information to model the normal pattern of input video samples.The framework cascade two adversarial Autoencoders:one that uses the U-net structure to be trained to learn spatial features;and the other that uses the improved U-net structure to be trained to learn temporal features.(3)We have improved the internal structure of the network.By introducing optical flow into the temporal information network,more temporal information is introduced to make the network more capable of time series modeling.By stacking the output of the spatial information network with the input of the time information network and then entering the temporal information network,the network can finally predict future frames.(4)The anomaly is detected by calculating the difference between the predicted future frame and the actual future frame,and the measurement of the difference is improved.Experiments on two data sets verify the effectiveness of our method.