Research On Anomaly Detection In Surveillance Videos Based Algorithm On Deep Neural Network

Anomaly detection in surveillance videos is the core of the intelligent monitoring system,which has high research value and application value in both academia and industry.However,the complexity of monitoring scenes and the diversity of behavioral events make video anomalous behavior detection a very challenging task.The two key steps that affect the accuracy of the abnormal behavior detection algorithm in the surveillance video are the description method of video behavior characteristics and the construction of the video behavior model.According to the different semantic information level of the video behavior description method,the characteristic description method of video behavior can be divided into target-level feature description method and pixel-level feature description method,according to whether the video abnormal behavior detection model uses deep neural network,the video abnormal behavior detection model is divided into detection model based on traditional machine learning algorithm and the detection model based on deep neural network.In recent years,deep learning has achieved very good results in the field of artificial intelligence,and methods of using a deep neural network models to detect abnormal behavior are emerging in the field of abnormal detection research.For simple and artificially specified abnormal behaviors,GANomaly,a semi-supervised generative adversarial network,is improved and applied to video abnormal behavior detection.Specifically,in order to make the network model more robust,the use of label smoothing can cause the network to produce tighter clustering and greater separation between categories;in order to prevent the model from collapsing and speed up the convergence of the model,in the confrontation training In order to make the appearance information of the image generated by the generated network more accurate,the reconstruction loss is divided into intensity loss and gradient loss.In abnormal behavior detection,the normalized reconstruction coding loss is used as the abnormal score to measure the abnormality of the video frame.Test and prove the stability and accuracy of this method on the public data set UCSD datasets containing category label information.In order to further make the abnormal behavior detection research more practical,for the specific and realistic violent behavior,the pseudo three-dimensional residual neural network P3 D Res Net was improved and applied to video abnormal behavior detection.Specifically,in order to overcome the problems of inter-class complexity and intra-class variability of human behavior,and the huge workload that would be brought by using accurate video frame-level tags,a multi-example learning method was used to learn video Behavior pattern.In order to optimize and improve the pseudo three-dimensional residual neural network,a 1×1×1 convolutional layer is added to the jump connection block structure,and a batch normalization operation is added after each convolutional layer,and the final classification layer is changed to three-Layer FC neural network.Each feature vector from the three-layer fully connected neural network is then output through the SVM of the output layer,and the abnormal score of the corresponding video segment can be obtained.The accuracy and robustness of this method were tested and proved on a large-scale public data set on violence,UCF-Crime.