Research On Video Abnormal Behavior Detection Method Based On Generative Adversarial Network

Abnormal behavior detection based on surveillance video is of great significance to social stability and public order.This paper classifies and analyzes the existing abnormal behavior detection algorithms,and conducts in-depth research on abnormal behavior detection based on generative adversarial networks.Generative Adversarial Networks perform well in abnormal behavior detection,but traditional Generative Adversarial Networks detection methods have problems such as low utilization of shallow spatiotemporal features,and gradients disappearing in deep network training.On the basis of Generative Adversarial Networks,this paper proposes a Two deep learning models based on prediction ideas,the main research contents are as follows:A variety of modal information,such as subject,subject behavior,scene,timing,etc.,needs to be considered for behavior analysis in video.Due to the powerful learning ability of deep learning models,methods such as generative adversarial networks are widely used in behavior detection tasks.Multi-feature extraction analysis and modeling of video.However,the detection of traditional generative adversarial networks has problems such as low utilization of shallow spatiotemporal features,and gradient disappearance due to deep network training.Under the framework of Generative Adversarial Networks(GAN),this paper proposes a generative model that incorporates a gated self-attention mechanism..It can suppress the expression of feature information in the input video sequence that is not related to the anomaly detection task,pay attention to the salient features of the video sequence related to the task,and assign weights to the shallow features and deep features in the model feature extraction process.The spatial features are fully extracted.At the same time,the lite Flownet optical flow network is used to focus on the temporal dimension of video sequences to ensure the continuity between video sequences.The experimental results of the model on the CUHK Avenue and UCSD datasets are higher than the typical methods of other classes on the AUC index,and this method achieves a higher effect on the accuracy of abnormal behavior determination.In order to further realize the extraction of spatial and temporal features of video,the second work of this paper proposes a generative adversarial network structure based on recurrent residual network from the perspective of network structure and convolution unit module.Specifically,it is mainly optimized for the generation network module,combined with the structural advantages of the recurrent network,the residual network and the U-net network,the recurrent structure is introduced into the convolutional layer of the U-net,and the convolutional unit of the backward sampling is used.Enter the residual unit,and fuse the input features and output features of this layer as the input of the next layer.Such a structure makes good use of the model to extract shallow features and ensures that the loss of feature information is minimized.In addition,the motion feature extraction of the video sequence is performed by using the Flownet network,and the optical flow prediction map is generated.Finally,the generated image and optical flow prediction are input into the discriminant network for result judgment,and the true and false results are obtained.Compared with the benchmark model,the AUC index and the quality of the generated images have been significantly improved,which reflects the effectiveness of the model in the use of features after the cyclic residual structure is introduced in the generation network part.Through the research on abnormal behavior detection algorithm,this paper found the insufficiency of spatial and temporal dimension feature extraction in video scenes,and then further researched and proposed two effective solutions to improve the accuracy of abnormal behavior detection.Contribute to the theoretical and experimental research of abnormal behavior detection in the future.