Correlation based search algorithms for motion estimation

Abstract

The measure of the 'goodness' of a motion estimation algorithm is governed by it's speed, the quality of motion compensation it provides, and the size of the resulting bitstream. Hence, algorithms should be evaluated based on this 'speed-quality-bitrate' tradeoff. Previously introduced fast motion estimation algorithms focus mainly on the speed vs. quality of motion compensation. In this paper, we introduce several new algorithms and evaluate them based on all three parameters. All these new algorithms exploit spatial correlation of motion vectors. These algorithms include a MAD (Mean Absolute Distortion) based spiral search, an Adaptive Window Size algorithm and two Majority Voting schemes. The algorithms are evaluated on several test sequences in the H.263 framework and the results obtained are very encouraging.