Background subtraction based on non-parametric model

In this work, a low memory and non-parametric based background subtraction algorithm (LMBS) is proposed by modeling the background model with a set of pixels. Instead of keeping a N-sized memory for each pixel model, a k-sized sample buffer is assigned to each pixel, and the background is modeled by 3?3?k neighborhood pixel values, which means the background model also contains space information. To adapt illumination changes, color metric based on YCrCb color space is used, which separates illumination information from chromatic information. A further processing considering pixel dynamics is conducted to adapt to geometry changes of background scene. The results on BMC (Background Models Challenge) dataset demonstrate LMBS algorithm performs as well as some widely used algorithms, with lower complexity.