An Iterative Approach for Segmenting Video Objects under Occlusion

In this paper we address the problem of segmenting foreground regions corresponding to a group of objects given their fragment features that were initialized before occlusion. The proposed approach includes foreground/ shadow segmentation and objects segmentation under occlusion. We present a weighted method to estimate shadows and foreground, by which we compute each pixel's features. Fragment features of objects in foreground were initialized before occlusion and updated by previous segmentation results. The problem is formulated in the framework of Conditional Random Fields (CRF), which is solved by using Gibbs sampling algorithm. To reduce iteration number, the Gibbs sampling algorithm was initialized by the mask that makes the normalized correlation of each object fragment between previous frame and current frame maximal.