Research On Crowd Abnormal Behavior Detection Method In Video Surveillance

Video monitoring system plays an important role in the field of public safety.The traditional video monitoring system is mainly realized by manpower,the workload is huge,low alarm correction-rate and latent-response.With the advancement of video surveillance network,highdefinition surveillance gradually has come into being,intelligent video surveillance marks the trend of the future.For a densely crowd which may include fleeing escape,fights,terrorist activities and other abnormal behavior,detection tactic is an important part of the field for intelligent monitoring,involving extracting and filtering the characteristics of the crowd in the video,and identifying the abnormal behavior and early warning.The ultimate purpose is to reduce the probability of disaster occurrence,and to ensure the safety security of public places.However,due to the presence of high crowd density,repetitive occlusion and complex behavior of people,the anomaly detection proves to be a challenging task in the related domain.Based on the problems mentioned above,the abnormal behavior detection algorithm this paper presents a new anomaly behavior detection algorithm by fusing both temporal and spatial features.The contents of research include saliency information feature extraction,crowd motion information feature extraction,crowd anomaly modeling technique and anomaly classification technology etc.The main work of this paper can be organized as follows:Firstly,in regard to saliency information feature extraction,the video frame is decomposed into image patches for local information extraction.In order to overcome the difficulty in combining the linear features in the existing models,the amplitude spectrum of the quaternion Fourier transform is applied to represent the color,intensity and orientation distribution of the image patches.The response mechanism of the human visual system is simulated,and the weight of the image block is considered by the human visual sensitivity.Based on the two factors and the saliency of variant scales,the final saliency map is obtained.The proposed method can effectively reduce the saliency of the background and improve the accuracy of detecting the crowd scene with complex background.Secondly,with respect to the crowd motion feature extraction,aiming at the problem with poor robustness found in fast information extraction,the global optimization rules is used in the grid optical flow estimation algorithm,which incorporates the dynamic social force model calculation to characterize the interaction of groups and individuals.To tackle the computation load and storage in redundancy,a new motion information descriptors is put forward,that is,the interaction force histogram,which effectively describes the crowd motion feature.Thirdly,in the video frames,the abnormal events can be described as a prompt change of features in the spatiotemporal domain.In this research,a fusion of saliency information and the social force model of unusual events detection algorithm is proposed,and the spatial temporal features will be extracted into the combination of support vector machine which is trained to classify the abnormal events.Finally,a given number of comprehensive and extensive tests are carried out on three different data sets.The results indicate that the proposed algorithm has satisfactory accuracy and robustness for the detection of abnormal events.