Hierarchical Estimation of Displacement Vectors in Image Sequences

Abstract

Abstract The problem of estimating the displacement vector for a moving object from a consecutive image sequence is one of the most important aspects of dynamic image processings in various fields such as image coding, remote sensing and robot vision. The displacement vector can be determined by the matching method, wherein the displacement vector is determined through the matching among subregions, or by the gradient method, wherein the displacement vector is estimated from the relation between the gray‐level gradient of the image and the gray‐level difference between images. Compared with the matching method, the gradient method has the advantage that the computation is simple and complex motion can be handled, although it is difficult to interpret global motion and the result is more vulnerable to noise. This paper proposes hierarchical estimation in which the advantage of the gradient method is utilized while coping with the demerits. The result is a more accurate estimation of the displacement vector. In the proposed method the hierarchical representation of the image is predetermined. The global motion estimated at a coarse level is transferred to more detailed levels of this hierarchy to compensate the previous estimation. By such a scheme, both the global and the detailed motions can be estimated with high accuracy and stability. By experiment it is demonstrated that the displacement vector can be estimated with high accuracy for image sequences for which estimation has been difficult by past methods.