Fast non-parametric background subtraction for infrared surveillance

Background subtraction is a method typically used to extract foreground objects in image sequences taken from static cameras by comparing each new frame to a background model, and it plays an important role in many vision application systems. In this paper, we introduce a non-parametric background subtraction method. Standard kernel density estimation method is very time consumptive, so it is modified by substituting the Gaussian kernel function with Epanechnikov kernel function and some optimizing techniques are adopted to improve its performance. As fluctuation is the intrinsic character of infrared image, we develop a bi-threshold updating method and a gradient based post-process method to reduce false positive error. Experiments show our method can extract intruding objects effectively and it outperforms threshold based method, especially when the intruder is not salient.