Fast and Effective Background Subtraction Based on ELBP

Background subtraction is an important component of many computer vision systems. In this paper, we present a fast and efficient texture-based method for background subtraction in a video sequence. At first, a novel LBP (Local Binary Pattern) operator called εLBP is applied to extract the local description of each pixel. Then, the kernel of background subtraction is performed based on two principle phases, i.e. the background building phase and the foreground detection phase. In the former phase, the probability that a pixel belongs to foreground is calculated based on the difference between the current εLBP and the adaptive mean εLBP. In the later phase, a thresholding operator is applied on the probability to determine whether a pixel can be classified as foreground, and the adaptive mean εLBP is updated by a user-settable learning rate. Finally, our approach is evaluated against several video sequences compared with the traditional MoG model. Experiments show that our method is suitable for various scenes and is appealing with respect to robustness.