Video Background Modeling Based On Optimization Algorithms Of Robust PCA

Video-based motion analysis mainly detect, track, classify, and recognize targets from the videosequences that contain a variety of moving targets, and all of the technology is based on thebackground modeling technique. Therefore, the study of background modeling has both importanttheory significance and application value.Unlike traditional methods, the video background modeling based on Robust PCA can obtainstable and precise background without training step previously, and it is robust to the change ofscene. This paper analyses the status and related application technologies of background modelingmethods at first, and introduces the foundational theory of robust PCA. Then it focus on thesubproblem about singular value decomposition, the ALM (augmented Lagrange multiplier)algorithm and ADM (alternating direction method) algorithm for video background modeling, andimprovement is carried out to achieve better results.The main computation of Robust PCAalgorithms is the singular value decomposition, and mostalgorithms only need to get the singular values that greater than a certain threshold. As the complexstructure and low computing efficiency of PROPACK, this paper proposes threshold-based lineartime singular value decomposition by improving the linear time singular value decompositionalgorithm, which is simple to understand with high efficiency.The ALM algorithm is very well for robust PCA, and it divided into exact ALM and inexactALM. As the inexact ALM algorithm is better for computing efficiency, the research bases on it inthis paper, removing PROPACK package while introducing the threshold-based linear time singularvalue decomposition. It can be found from the background modeling experimental results that theimproved algorithm has better modeling result and computing efficiency.The ADM algorithm is also a good method for robust PCA, and it has good effect when doingexperiments based on synthetic data. Thus, the paper makes it for the real video backgroundmodeling, and it can be found that the result of the experiment is varied when changing the value ofthe parameter. With a proper and smaller, the ADM algorithm will have a good backgroundmodeling effect, but it still need to be improved when compared with ALM algorithm.