Non-incremenal 3D Reconstruction Based On Multiple View Geometry

3D reconstruction is the premise and foundation of understanding and analyzing the world using cameras in compute vision. Compared with reconstruction using depth camera and 3D scanner, reconstruction based on multiple view geometry is much more flexible and cheap. Due to these reasons, 3D reconstruction based on multiple view geometry has become a hot spot in the field of compute vision.Simultaneously, non-incremental 3D reconstruction base on multiple view geometry is faster and more accurate compared to traditional incremental methods. Therefore, it has become a focus in the field of multiple view geometry reconstruction. The core idea of non-incremental 3D reconstruction is to recover the global position of every camera using pairwise relations. However, not all of the pairwise relations are correct, so the outlier of the pairwise relations must be removed before computing the global position of every camera, and this is one of the difficulties. Another difficulty is the computation of the camera translations because the relative translation is usually in the different scales and with low accuracy. On this basis, the main work and achievement of this dissertation are as follows.(1) On the base of analysis and researches on the existing methods of outlier removal of the pairwise relations, a new method of outlier removal is proposed in this dissertation based on AC-RANSAC. Experiments on different datasets have shown that the proposed method is more stable, which is because this method successfully avoid setting an unreliable threshold.(2) In order to improve the accuracy of the computation of relative translations, a method that combined golden section with linear programming is also proposed in this dissertation to compute the triple view translations. Experiments have shown that the proposed method performs better than other methods.(3) The core models of an integral 3D reconstruction system is also developed in this dissertation including camera calibration, feature extracting and matching, sparse reconstruction, dense reconstruction and texture mapping. For the model of sparse reconstruction, we fuse the methods proposed in this paper with the open source SfM system OpenMVG. Meanwhile OpenMVG has also been modified to handle images with different focal lengths.