Predictive motion estimation with global motion predictor

Predictive motion estimation techniques have become widely used in video coding systems due to their reasonable coding performance and low computational requirement compared to the brute force full search method. However, most existing predictive motion estimation algorithms focus mainly on local motions and consider motion predictions from only neighboring image blocks (either spatially or temporally). In this paper, a reliable global motion predictor or GMP is introduced to improve predictive motion estimation for sequences with fast camera motions. A simple global motion model is applied for global motion estimation. The derived global motion vector for each macroblock is then used as the GMP during the predictive motion estimation for all possible block modes or types, together with up to 8 local motion predictors for P frame motion estimation and up to 12 local motion predictors for B frame motion estimation. The comparison between the coding performance with and without GMP shows very significant differences; roughly 1-2 dB’s improvement is achieved by GMP for sequences with very fast camera motions. The computational overhead for GMP is very small, less than 2% of the total computation cycles in our H.264 encoder.