A Method of 3D Reconstruction from Image Sequence

In this paper, a method of 3D reconstruction from an image sequence acquired by a moving camera is presented. The internal parameters and the motion of the camera are absolutely unknown. Firstly, the features of the reconstruction object are detected in each image, and are matched between image pairs. A new method of angel filtering is used to eliminate false correspondences. Fundamental matrix can be computed and projective reconstruction is set up. Secondly, camera self-calibration method is used to get the internal parameters. Epipolar geometry is used to obtain a pairwise image rectification. Dense correspondence matches are then computed. Finally, the 3D model is reconstructed based on the self-calibrated result and dense correspondences map. The feasibility of this approach is illustrated by the experiment.