Performance analysis of Lab2000HL color space for background subtraction

Abstract

Background subtraction techniques are commonly used to identify moving objects in computer vision applications. This is still a challenging problem, especially when there is a non-stationary background such as a waving sea, or in the case where camera oscillations exist, or when videos have non-stationary backgrounds because of sudden changes in lightning. One of the most significant sub-tasks of a generic background subtraction technique is the background modeling step, which determines how background will be represented. A wide range of the literature is upon development of statistical models for background modeling. Especially, Gaussian Mixture Model (GMM) is a basic method. In this method, values of each pixel's features with respect to time are represented with a few normal distributions. The problem with which features pixels will be represented is an important research topic. Recent studies involve applications using different color space with both pixel and region based features. In this study, in addition to color spaces used in literature, new color space which have linear hue band and named as Lab2000HL is aimed to test. The segmentation of foreground/background performance is measured with average precision rate. Nine different videos from I2R dataset having non-static background examples are used as test dataset.