Background Modeling Based On Local Low-rank Robust Principal Component Analysis

Background modeling is an important research area and the foundation of manyquestions in computer vision. As large increase in captured video data, building thebackground model for complex videos or image sequences in a short time has becomea popular research topic.Most of traditional background modeling methods are based on a statistical modelin pixel level, which the parameters of models is learned in the training phase andcontinuously updated in the background modeling process. In recent years, manybackground modeling methods based on robust principal component analysis havebeen proposed. Those methods have good background modeling results even whenthe light conditions suddenly changes, and also don’t need a training phase, whichmake the related research become a hot research direction in background modeling.A prevalent assumption in robust principal component analysis is that theobserved matrix is composed of a low-rank matrix and a sparse matrix, or, theobserved matrix is of low rank. This paper presents a new hypothesis that theobserved matrix is locally of low-rank (instead of globally of low-rank). Specifically,the observed matrix can be represented by several independent parts which are of lowrank respectively. Under this assumption, we design the method of backgroundmodeling based on local low-rank robust principal component analysis, which dividethe observed matrix into several parts, and process each parts separately with robustprincipal component analysis, and combine the results for each part together.Experiments show that our method has better results than exist global robust principalcomponent analysis under circumstances which the background is significantlychanged. Moreover, our method is faster than the exist global robust principalcomponent analysis method, and is adapted to parallel computing, which make ourmethod could better satisfy the speed requirements of real-time background modeling.