Abstract: To analyze quickly the motion information of videos, a fast optical flow estimation algorithm for compressed domain is proposed. Firstly the spatial partial derivatives of image luminance are estimated by using two AC coefficients, and then the predictive residual errors and motion vectors of blocks are appended to estimate the temporal partial derivatives. In addition, after the detailed analysis of those macro-blocks that had no forward motion estimation, the motion information is given approximatively relative to their forward reference frames. And then, a amendatory partial derivative estimation is given for the inconsecutive image blocks when occlusion and cut had happened. Finally, optical flow estimation is performed based on the least square method and partial derivative estimation. Experiments indicated that this method can more accurately estimate the optical flow than the L-K method in pixel domain and the exist method in compressed domain. Moreover the proposed method can greatly reduce the compute time than the estimation in pixel domain.