Low complexity motion estimation algorithm by multiresolution search for long-term memory motion compensation

We propose a new low complexity motion estimation (ME) algorithm using multiresolution motion search for long term memory motion compensation (LTMC). While multiresolution motion search has been used for standard single-frame motion compensation, here we introduce several novel techniques to exploit it efficiently in the context of LTMC. The proposed algorithm compute a coarse motion vector from a lower resolution sequence to locate motion search windows (MSW). Because the coarse motion vectors are good approximations to the actual vectors, small MSW whose sizes are chosen based on the error surface smoothness information can be used without significant loss in quality, leading to significant reduction in complexity. Our approach also incorporates temporal reduction of the motion search range to reduce the complexity further. The proposed algorithm provides particularly significant gains for video sequences which contain large motions, where directed search also results in promising gains in PSNR over non-directed search.