Research On Moving Object Detection Based On Robust Principal Component Analysis

Moving object detection is a cornerstone task of computer vision,which provides guarantee for the completion of other advanced tasks in the follow-up.It has a wide range of applications in many fields such as vehicle navigation,pedestrian detection,and intelligent transportation.The complexity of real scenes directly affects the effect of moving object detection,and its research is of great significance.Moving object detection methods based on robust principal component analysis(RPCA)have received widespread attention.This dissertation conducts research on moving object detection based on nonconvex robust principal component analysis.The main innovations are given as follows:(1)Aiming at the problem of large error caused by low-rank matrix using the nuclear norm approximation when classical robust principal component analysis is used for moving object detection,a moving object detection based on nonconvex motion-assisted robust principal component analysis(NMARPCA)method is proposed.This method uses nonconvex?norm to approximate the low-rank matrix.It also considers that the background is still sparse in the discrete cosine transform domain and uses the ₁l norm to represent its sparsity.In addition,it also introduces a random matrix that satisfies a uniform distribution,and the elements in the matrix are expressed as the probability of background pixels;And then,the alternating direction method of multipliers(ADMM)is used to sove the proposed NMARPCA problem.Finally,the proposed NMARPCA method is used to detect the moving objects of 7 original videos and their corresponding videos with Gaussian noise in CDnet and I2R datasets.The experimental results show that the proposed method is superior to many other moving object detection methods based on RPCA.(2)In order to reduce the negative impact of complex background on moving object detection in real scenes,and solve the problems that ₁l norm can not better approximate the sparsity function of the traditional robust principal component analysis model,as well as need to manually adjust the trade-off parameter,a moving object detection based on nonconvex norm and Laplacian scale mixture with motion map(NNLSMM)method is proposed.This method uses nonconvex?norm and the Laplacian scale mixture(LSM)to approximate the low-rank matrix and sparse matrix in the traditional robust principal component analysis model respectively,and uses the optical flow method to obtain the dense motion field matrix of each frame of image data,so as to further improve the accuracy of moving object detection.At the same time,the alternating direction method of multipliers is used to solve the proposed NNLSMM problem.Finally,the proposed NNLSMM method is used to detect the moving objects of 7 original videos and their corresponding videos with Gaussian noise in CDnet and I2R datasets,and the experimental results verify the proposed NNLSMM method is more superior than some other moving object detection methods based on RPCA.(3)Aiming at the problem that the classical robust principal component analysis model does not fully consider the structural information of the moving object,and nuclear norm and ₁l norm cannot better approximate the rank and sparsity functions of this model,a moving object detection based on nonconvex truncated norm and structured Gaussian scale mixture(NTNSGSM)method is proposed.This method uses the truncated nonconvex?norm to approximate the rank function,at the same time considers the structured characteristics of the moving object,and uses the structured Gaussian scale mixture(SGSM)to better approximate the sparse matrix.And then,the ADMM is used to solve the proposed NTNSGSM problem.Finally,the proposed NTNSGSM method is used to detect the moving objects of several original videos and their corresponding videos with Gaussian noise in CDnet and I2R datasets.Through the analysis of the experimental results,the effectiveness and superiority of the NTNSGSM method are proved.