Motion estimation based on eigenvalue algorithm matching local image features

摘要

We propose a method for motion estimation based on the eigenvalue gradient method with regard for local image features. In the proposed method, motion estimation is first performed by the eigenvalue gradient method, and low-reliability areas are classified by image features. The estimation accuracy of such areas is improved by using two eigenvalue gradient algorithms. Blocks with multiple motions are processed with robust estimation. On the other hand, blocks with only one gradient are estimated by using a supplementary vector; in particular, the second eigenvector is replaced by the eigenvector of another block from the image area with a common edge. Experimental results show that the proposed method offers improvement of estimation accuracy at a small increase of computational cost.