Research On Online Video Surveillance Based On Vector Autoregressive Model

Daily safety monitoring in industrial areas such as petrochemicals and electrical is vital to corporate development and national security.For remote unmanned areas where pipelines and equipment are placed,traditional video surveillance methods are not only unable to detect screen anomalies in a timely manner,but are also very labour intensive.In order to quickly and accurately detect screen anomalies,realise online monitoring of video data in unmanned scenarios and help enterprises reduce their daily management costs,this paper accomplishes the following.(1)The video image data is a high-dimensional time series,and the cross-correlation and autocorrelation features of the high-dimensional time series are fitted by the VAR(1)model,and the coefficient matrix decomposition is innovatively introduced as a lowrank matrix and a sparse matrix method,i.e.,the least squares method is used to estimate the objective function of the parameters by adding the first and kernel parametrization to achieve the coefficient matrix decomposition,and the sparse component and low-rank component are obtained,and the sparse component can capture the changes of the pixel points in the fault region in the image.(2)Applying the control chart to the direction of video monitoring,using the advantage that the classical EWMA control chart is sensitive to small offsets,constructing the monitoring statistics in the EWMA control chart based on the sparse matrix,and according to the two-stage method of statistical process control,firstly searching for the control limits in the controlled state using the Monte Carlo simulation method,and then realising real-time monitoring of video data by comparing the monitoring statistics with the size of the control limits.In this paper,the improved method of constructing the monitoring statistics based on sparse components is used to conduct experiments,and it is compared with the two methods of constructing the monitoring statistics based on transfer matrices changes and pixel differences,and the average running chain length of the three methods is compared in the example monitoring data.The results show that the VAR model fitting method with a matrix of decomposition coefficients is able to detect anomalies more quickly and monitor them more effectively.