High performance algorithms for MPEG motion estimation

Abstract

In this paper, we propose two high performance algorithms for the motion estimation searching. The first algorithm express minimum mean square error motion estimation in terms of a cross-correlation computation; this approach reduces the number of multiply-adds needed by almost factor of two, with further improvement possible when fast cross-correlation algorithms are used. In the second algorithm we express the motion estimation computation in a form similar to the matrix-matrix multiplication computation. We propose an efficient algorithm for this computation. The algorithm exploits the memory hierarchy of the target processor and gives optimal performance for this formulation. In certain cases, both approaches can be used simultaneously.