As traditional single sequence scene reconstruction method is easy to loss scene information and has a lower data utilization, this paper proposes an approach to the registration between multiple sequences for scene reconstruction. The method uses a similarity transformational as the registration model and proposes the separating initial registration method to deal with the scale, rotation and translation respectively. Then, it fits the observed plane to suppress the noise data and select the co-visible points. Finally, the registration between multiple sequences with different scales, positions is performed and the registration results are applied in successive subsequent reconstruction. The experiment shows that the initial error has been reduced from 17.69% to 46.86%, and the terminated error reduced from 27.5% to 71.96% after suppression of noise data along the oriented orientation. Moreover, the 10 596 vertices are reconstructed from 47 frame images, more better than 3 893 vertices reconstructed by traditional single sequential method. The proposed method full makes use of the observation images and allows the ultimate fitting surface of the scene to have more scene details.