Segment Based Diffusion - A Post-Processing Step (Not Only) for Background Subtraction

A classical approach to video object segmentation is background subtraction. Background subtraction starts by estimating a model of the background image of a scene and then calculating the likeliness that a given pixel of the current camera image belongs to the background model. Typically this is done by subtracting the background image from a given frame, where the difference image is usually thresholded and post-processed to find object boundaries. In this paper we present a method for enhanced post-processing that exploits color and texture information of the original video frame. This way we are able to extract pixel-exact object boundaries. Based on direct color segmentation of the video frame, an iterative method analog to biological diffusion and physical heat transfer processes, spreads information from the difference image over segment boundaries. For this purpose, diffusion resistances are defined between adjacent segments, based on color and texture similarities and common contour length. An iterative process calculates and transfers the flux of 'difference energy' between segments of the difference image. The resulting image allows for easy segmentation by thresholding. Experimental results show the validity of our approach.