Video Anomaly Analysis Based On Graph Convolutional Network And Anomaly Unbalance Suppression

With the development of video technology and the improvement of people's awareness of safe life,surveillance video analysis technology is getting more and more attention.Video anomaly analysis,as one of the key technologies of surveillance analysis,has been a hot spot for research,and widely applied in the fields of road traffic warning system,community security monitoring system,surveillance forensic search,etc.There are complex scenes and long time series characteristics in surveillance video,the probability of anomalous events in the video is low and the probability of different abnormal events is also different,so how to quickly discover and classify anomalous events based on long time series,complex environment and imbalance of anomalous categories has become the breakthrough point of research,which is also a difficult point at present.Based on a large amount of foreign research literature,firstly,the current research status,challenges and problems of video anomaly analysis are described in this thesis,and the application areas are also briefly described.Secondly,a detailed introduction is given for video analysis techniques,graph neural networks of associative learning and weakly supervised and unsupervised methods for video anomaly detection.This thesis takes the weakly supervised video anomaly analysis as the research point,combines the spatio-temporal correlation embodied in the anomalous events in the video,and proposes the learning method based on spatio-temporal fusion graph network to enhance the localization ability of the model.On the basis of better localization ability,this thesis considers the sparseness and imbalance of the anomaly,and the method based on salient attention model and class imbalance suppression is proposed to improve the analysis ability of the model for the anomaly.Based on the above methods,a prototype system for video anomaly analysis is developed.The main research contents of this thesis are as follows:(1)Considering the strong spatio-temporal correlation characteristics of the occurrence of anomalous events in videos,a video anomalous event detection method based on spatio-temporal fusion graph network learning is proposed.Utilizing the temporal and spatial relationships of each video segment in the anomalous video,the video segment features are embedded into the nodes in the graph.The spatial similarity graph is constructed with Top-k similarity and the temporal continuity graph is constructed with Top-k continuity.The form of graph structure is used to express the inner connection of the video.To enhance this connection even more,the temporal continuous graph and the spatial similarity graph are fused by adaptively weighting to form a spatio-temporal fusion graph convolutional network,and learn to generate video features with stronger robustness.Combining with the sparse constraint term of the graph to reduce the oversmoothing effect in the graph model and improve the final detection performance.The experimental results on three datasets verify the effectiveness of the spatio-temporal fusion graph network model,which can effectively improve the detection performance of video anomalous events when compared with frontier methods.(2)Considering the sparsity of anomalies in videos and the imbalance of anomaly classes,a weakly supervised video anomaly analysis method is proposed which based on salient attention model and class imbalance suppression.In order to solve the sparse expression of anomalies in video,attention model and combined with top predicting ratio sifting method are used,which ensures the sparseness of anomalies and maintains the purity of anomaly information.The anomaly features are more conducive to classification after screening and aggregation by GCN method.In the classification phase,the anomaly imbalance loss is applied to reduce the impact of the loss function percentage brought by the imbalance category,so that the model can learn more information of different anomaly categories and enhance the classification ability of the model.A large number of experiments are conducted on the common video anomaly analysis dataset,and the experimental results show that the model can more accurately locate the anomaly interval and give a more accurate category.(3)A video anomaly analysis prototype system is developed using Python,QT interface development tool and corresponding basic framework,which integrates three major functions: video feature extraction,model training and video anomaly analysis.There are three characteristics for the system: beautiful interface,easy operation and good interaction experience,which can satisfy the user's needs for video analysis in daily surveillance.