Eigen-Patch Based Background Subtraction

This paper describes a method for modelling the background of a static video camera as a combination of intensity and texture. Each square patch of the image is described as a uniform component, representing the intensity, and a weighted combination of zero mean eigen-patches which represent the texture. The intersection of the weight distributions provides a measure of the distance between two textures. A standard multi-mode background representation for the background is then constructed, to produce a map for the distance from the background of each pixel in every new image frame. This distance map is smoothed using a cross-bilateral filter, and thresholded to delineate foreground objects. Some tests using the nine video sequences from the I2R dataset are used to compare the proposed method with some recent techniques.