Abstract: A fast and memory-saving registration algorithm is proposed for image sequence-based computer vision applications. First, a new scheme, which is independent of the image content, is adopted to generate a series of feature points. Then, a forward-and-backward tracking approach is used to obtain the reliable feature point pairs. In the approach, the forward tracking is performed to get all the potential feature point pairs, and the backward tracking is utilized to measure the forward-backward discrepancies which enable selection of the final reliable point pairs. Finally, building upon these reliable pairs, the homography is computed by using normalized direct linear transformation. Due to the intensive exploitation of the interframe continuity existing in image sequences, the proposed registration algorithm provides comparable experimental results to current state-of-the-art techniques, while using a fraction of the computation time and a fraction of the memory as well. Specifically, with a memory usage of 421 kB only, this algorithm runs at 32 frames per second for a sequence with an image resolution of 480360.