Background updating with the use of intrinsic curves

Abstract

A primary tool to extract information about moving objects is background subtraction. In this technique, the difference between a model of what is static, or background, and the current image of the scene gives information about what is in the prime plane or foreground. This study focus on the pixelwise updating mechanism of the background model throughout the analysis of the images provided by a fixed camera. The concept of intrinsic curves, early introduced in the field of stereovision, is extrapolated to the problem of detecting the moving boundaries. We use a mixture of Gaussians to register information about the recent history of the pixel dynamics. Our method improves this model in two ways. Firstly, it reduces the chances of feeding the mixture of Gaussians with foreground pixels. Secondly, it takes into account not just the scalar pixel value but a richer description of the pixel’s dynamics that carries information about the interpixel variation. Ample experimental results in a wide range of environments, including indoors, outdoors, for a different set of illumination conditions both natural and artificial are shown.