The Research On Surveillance Video Anomaly Detection Technology

With increasing awareness of public safety, the technology of video surveillance has been widely used, in which the crowd anomaly detection has been a challenge. Machine learning method was applied in video anomaly detection system to improve the efficiency of the system. Anomaly detection in surveillance video scenes is technology of public safety,which has become a hot research at home and abroad.By analyzing the existing problems of the abnormal detection in the intelligent video surveillance system, the paper has researched the abnormal detection based on the analysis of the video features information and the population models.Meanwhile,this paper also explores the abnormal detection framework of distributed video,which is based on the cloud platform.The main work is as follows:First, this paper analysises the characteristics of the crowd directly to detect the abnormal scenes. In order to improve the accuracy rate and reduce the detection time of abnormal detection, this paper proposes a hybrid optimization algorithm of feature selection and SVM training model. In this method, it will extract the feature datas and ruduce the dimension firstly, then build the SVM classification models. Feature selection and training model are interaction processes, while the feature selection and SVM model parameters were optimized simultaneously. The hybrid optimization algorithm based on inproved GA, which design adaptive cross and mutation, and integrate with SA. Experiments show that this method can quickly find the optimal subset of features and optimal parameters of SVM model. The accuracy of the video anomaly detection has been improved with the reducing the detection time in video abnormal detection.Secondly, this paper proposed PSO optimized social force model to accurately simulate crowd movement state. This method describes the speed of pedestrian in SFM by optical flow, then takes PSO to improve the traditional SFM modeling. The social force vector was inputed to SVM classifier, and prediction abnormal behaviors in the crowd. In particle advection method, the particles move to the area with smaller interaction force to simulate crowd behavior. Experiments show that this method can effectively improve the recall rate and precisio of anomaly detection analysis.Finally, in order, this paper explores the parallel anomaly detection framework based on Hadoop to deal with the mass surveillance video data. The framework study Map/Reduce key value and I/O format of video data, and then designs SVM classification algorithm based on Mapreduce. Experimental results show that, compared with the single node, video anomaly detection method based on Hadoop clusters at the same time to ensure the accuracy and dramatically reduce the detection time for video data.