Fast variable block-size motion estimation is a key issue for real-time applications of the H.264, whereas the subpixel refinement takes up much computational time as compared to integer-pixel motion estimation. We propose a new fast subpixel precision variable block-size motion-estimation scheme. This algorithm uses the statistical information, which comes from the motion activities of the macroblocks (MBs) in the previous frame, to predict the characteristics of MBs in the current frame. Additionally, the distortion values and motion vectors of MBs in the previous frame are also considered as prior knowledge, based on which we can make decisions on early mode selection and early termination, and on whether or not to skip some candidate modes and candidate checking points. The intermediate results of subpixel motion estimation are used together with the prior knowledge to reduce subpixel search time when searching for stationary blocks. Our new directional information strategy is used in both integer-pixel motion estimation and subpixel motion estimation to accelerate the search procedure. Moreover, our algorithm can eliminate the subpixel motion estimation of all the unselected subpartition modes. The computational resources can then be spent on the modes and locations that deserve to be searched more than others. Extensive experimental is been done, the results of which show that the speed of our approach is nearly five times that of the fast algorithms in H.264 JM, with a better peak signal-to-noise ratio and better bit performance.