Motion-based background modeling for foreground segmentation

Abstract

Foreground segmentation is an elementary process in an intelligent visual surveillance system.When a background is changed dynamically, it is difficult to distinguish the background from the foreground. In this paper, we propose an unified statistical framework of foreground segmentation and motion estimation method. In this method, a motion prior distribution is recursively estimated by a particle filter model. The motion prior probability plays a key role as a weight of each pixelwise observation model in classifying a pixel as background or foreground. We use a kernel density estimation method for the observation model. Using a temporal diffusion kernel, we emphasize recent observations. A soft selective updating rule is also suggested and this rule can overcome a deadlock problem.The algorithm can be applied to images acquired from a fixed camera. Experimental results with many real image sequences showed the validity of our method.