Robust video segmentation employing variable bandwidth object boundary

A pixel-wise accurate and interactive video segmentation technique is proposed in this paper to generate a binary mask of moving objects optimized for MPEG-4 object-based video coding. The initial object mask is tracked temporally, and regions are classified into the background region, the object region, and the uncertain region where the pixel is within a certain distance from the tracked object boundary. We exploit the color feature in the L*u*v* space, and use a non-parametric gradient-based iterative color clustering algorithm, known as the mean shift algorithm, to generate the region marker in the background and the object regions. Furthermore, we use the watershed technique to classify pixels in the uncertain region into the background and the object. Finally, by employing a variable bandwidth region boundary, the final mask for moving objects is optimized for MPEG-4 shape coding efficiency. The shape coding efficiency can be improved up to 30% with negligible loss of perceptual quality