Method for encoding a video signal using feature point based motion

Abstract

A method for providing an accurately predicted current frame, comprising the steps of: selecting feature points in the previous frame and detecting a first set of motion vectors for the feature points; defining quadrangles by connecting the feature points; defining horizontally split triangles by splitting quadrangles and calculating a variance for each triangle and comparing the variance to select the smallest variance as first variance; defining vertically split triangles by splitting the quadrangles and calculating the variance for each triangle and comparing the variance to select the smallest variance value as second variance value; comparing the first and the second variance and selecting the smaller variance as a third variance and providing triangle information representing the split direction for the third variance; determining quasi-feature points on the current frame and a second set of motion vectors for the quasi-feature points; defining quadrangles formed by connecting the quasi-feature points, and splitting each quadrangle into two triangles; determining a predicted position on the previous frame for pixel contained in triangle of the current frame; determining a third set of motion vectors for the pixels contained in triangle of the current frame; and providing value of each pixel to be contained in the predicted current frame by using the second and the third set of motion vectors.