Motion Estimation and Region Segmentation via Functional Optimization

High performance coding of image sequences requires efficient motion estimation and compensation. Block matching techniques are the usual choice. However they produce irregular motion fields and require subsequent regularization. This paper reports on a motion estimation technique that merges block matching with polynomial approximation and region growing to achieve such a regularization of the measured motion field. As a byproduct, a segmentation of the scene into regions with smoothly varying motion is obtained. Experimental results show that the obtained motion fields, though much more regular than the ones produced by simple block matching, still are very effective in giving a small prediction error.