To reduce the complexity of AVS-M sub-pixel motion vector search，this paper proposes a sub-pixel motion estimation fast searching algorithm based on prediction．The algorithm expels the less possibility points by doing matching error relation of sub-pixel points selected analysis

which assures the accuracy of sub-pixel search. Through effective search strategy

compared with the half-pixel full search algorithm and CBFPS

experiments show that the proposed algorithm can respectively reduce the sub-pixel search points by 65.25％87.69％ and 43.67％67.38％ on average PSNR lost less than 0.01dB to the video sequences with different motion characteristics. The algorithm efficiently decreases the computational number of sub-pixel motion estimation．