Video retrieval is a current hot research area. Most of the past algorithms are done in pixel domain, which need many decode calculations. What is more, the same matching algorithm is used to all the video clips in the same way, which wasts many unnecessary calculations. A new hierarchical retrieval method based on example video for MPEG video is proposed: firstly the dct_dc_size field in I frames is used to locate the suspected videos quickly, then the retrieval result scope can be further reduced by analyzing the spatiotemporal distribution of motion vector in B frames using tomography method, at last DC images precise matching analysis is used to validate the retrieval result. The experimental results show that this method needs a few calculations and has higher precision ratio.