Video Anomaly Detection Base On Multiscale Transformation Of Graph Structure

In recent years,with the continuous improvement of people's living standard and the increase of population urbanization rate,public safety issues have become more and more prominent.In the monitoring scene,anomalous events can be promptly detected and actively rescued.It is of great significance of reducing the mass casualties and property losses of the masses.Aimed at the problems of large amount of data and complicated data analysis and processing in the video surveillance scene,the video surveillance system needs to be continuously improved,and has become more intelligent and efficient.In this paper,we mainly present the algorithm of video anomaly detection with multi-scale transforms into graph structure and the graph structure of video features in multivariate logarithm Gaussian distribution.The detailed contents are as follows:This chapter presents a method of video anomaly detection with multi-scale transformation of graph structure.In view of the fact that the video feature data onto the monitoring scene exists on the vertexes of the weighted network space structure and the irregular network graph structure can better express the spatial relationship of the features,the paper describes the feature data in the video by using the graph structure.Under the premise of preserving the spatial structure of the video features,we first construct the network structure of the optical flow features and make use of the iterative scaling transformation of the graph structure of optical flow features of the relevant constraints.This will effectively reduce the number of the optical flow features in the video,thereby increasing the computational speed of video anomaly detection.The scale transformation of the graph structure of optical flow features is to choose the vertices by using the polarity of the largest eigenvector corresponding to the graph Laplacian of the graph structure of optical flow features to complete the graph down sampling operation,then we use the kron reduction to construct the intrinsic connections between the vertices.In this paper,anomaly detection has been carried out under the UMN dataset and the more challenging WEB dataset.The experimental results show that the proposed algorithm can greatly improve the anomaly detection with slightly lower detection accuracy.In this chapter,we propose a graph structure reduction of video features in multivariate logarithm Gaussian distribution.For the graph structure of optical flow,the phase information on the optical flow eigenvector at the vertices is contained in the optical flow field.The amplitude of the optical flow eigenvector reflects the change of the target feature of the continuous video frame.Since the amplitude of the optical flow eigenvector are all greater than zero,the amplitude of the optical flow eigenvector at the vertices of the graph structure in the video scene obeyed the multivariate logarithm Gaussian distribution.Under this premise,this chapter combined with the down sampling method of graph Laplacian maximum eigenvector and the down sampling method of the amplitude mean of the optical flow eigenvector in the graph structures to fit the multivariate logarithm Gaussian distribution,So that we can make the reduction in the graph structure of optical flow features and achieve the purpose of feature optimization.In order to verify the practical effect of the method,we experimentally verify the video anomaly detection on the UMN dataset.Experimental results show that this method can effectively reduce the amount of data and can achieve rapid anomaly detection.