The direct mode coding in the bi-predictive pictures (B-pictures) can efficiently improve the performance of bi-predictive coding, because it exploits the temporal correlation by bi-directional prediction from both forward and backward reference pictures, and meanwhile it does not require any bits for coding the motion vectors. Accordingly, how to accurately calculate the motion vectors in terms of direct mode coding is very important to obtain better prediction values. In other words, it is more desirable to obtain the true motion vectors, because the derived motion vectors need not be coded. The traditional direct mode coding usually derives the motion vector of the current block by scaling the motion vector of the co-located block in the subsequent reference pictures. It is based on the assumption that the near blocks would have the same motion trajectory. In this paper, we present an improved method based on combining motion vector tracking technique and spatial motion vector prediction technique. The proposed method can accurately calculate the motion vectors for the direct mode coding.