Research On Abnormal Behavior Recognition In Video Monitoring

Video abnormal behavior recognition is one of the important functions in the intelligent monitoring system,which refers to the use of related image processing and video processing technologies to automatically identify such abnormal behaviors,violent activities,and non-compliant behaviors in the surveillance video.However,in the actual application process,due to the limitations such as the ambiguity of the definition of abnormal behavior,the diversity of monitoring scenes,the complexity of the background,and the real-time nature,video abnormal behavior recognition faces huge challenges.At present,there are many studies using different methods to improve the accuracy and stability of video abnormal behavior recognition.In general,these algorithms can be divided into two categories based on traditional ideas and deep learning ideas.However,the advantages and disadvantages of the existing algorithms are obvious,and there are always shortcomings in recognition accuracy and stability.This article first studies the traditional video feature extraction method.The biggest problem with traditional manual feature extraction methods is that the descriptors can't match the rich spatiotemporal features in the video.Convolutional sparse coding can perform global dictionary decomposition of pictures,and retain the connection between adjacent image blocks,so that it is expected to solve the problem of insufficient expression ability in feature extraction.Starting with global video anomaly detection,this paper proposes an anomaly detection algorithm based on convolutional sparse coding.On this basis,this paper combines the convolutional sparse coding with the local feature extraction method Histograms of Oriented Gradients(HOG)to complete the good combination of global features and local features in the video and improve the recognition of local abnormal behavior Accuracy and stability.It is easy to ignore the feature of time dimension in video is one of the main problems based on deep learning algorithm.To solve this problem,a 3D convolutional neural network is applied to extract the spatiotemporal features in the video,and adds optical flow features as constraints.The advantages and disadvantages of the reconstruction and prediction network based on deep learning are significant.The reconstruction network has high requirements for the diversity of training samples.The prediction network is often sensitive to time-domain anomalies and unstable to air-space anomalies.In order to solve this problem,we combine the reconstruction network and the prediction network skillfully.In the proposed model,the auto encoder reconstruction model and the prediction model are two parallel branches that share the encoder part,while giving full play to the advantages of the two,the stability of the network is improved.This paper uses three data sets(UMN,UCSD,CUHK Avenue)commonly used in video abnormal behavior recognition to evaluate the proposed global and local anomaly detection algorithms.In the UMN data set,a global anomaly detection algorithm based on convolutional sparse coding can accurately identify abnormal behaviors in different scenarios,achieving a recognition accuracy of 100%.In the UCSD and CUHK Avenue datasets,the effectiveness of the proposed local video anomaly detection algorithm is proved.The proposed algorithm has good recognition accuracy in different subsets.In the ped2 subset of the UCSD dataset,the proposed video anomaly detection algorithm of convolution auto encoder reconstruction and prediction network has reached 92.5% recognition accuracy rate.Experimental results verify the effectiveness of the proposed algorithm.In order to further prove the effectiveness of the proposed algorithm,this paper also carried out an anomaly detection experiment on actual video,the experimental results prove the effectiveness of the proposed algorithm.