Precise estimation of motion vectors and its application to MPEG video retrieval

We propose a new method to improve the correctness of motion vectors in MPEG video sequences. This method is based on our previous research of reliability metrics for motion vectors. Then, based on reliable motion information, we develop a motion-based hierarchical searching/matching model for video retrieval systems. The experimental results show that this model provides a good tradeoff between computational cost and the quality of retrieved data, which still remains a problematic issue in video information retrieval. This method promises to be an important technique for building efficient video retrieval systems