基于最优RANSAC算法的非增加式多视图三维重建

Non-sequential multi-view 3D reconstruction base on ac-ransac

非增加式的多视图三维重建算法不依赖于初始重建,而是首先求解全部的摄像机参数,然后一次性重建所有三维点,其难点在于如何精确的求得全局旋转矩阵和全局平移向量.针对摄像机旋转矩阵计算的问题,引入最优RANSAC算法来稳定剔除错误的二视图关系,避免了人为设定阈值的不可靠性,提高了全局旋转矩阵的计算精度.针对摄像机平移向量计算的问题,结合了黄金分割方法与线性规划来精确计算三视图相对平移向量,提高了相对平移向量计算的精度,从而提高了最终全局平移向量的精度.实验结果表明,设计的算法能有效的提高三维重建精度.

Non-sequential multi-view 3D reconstruction algorithm does not depend on the initial reconstruction.Firstly,the compute camera matrix of all views are solved,then all 3D points are reconstructed at once.The difficulty of non-sequential method is to compute the global rotations and global translations accurately.As to the computation of the global rotations,a robust algorithm is designed to remove the outliers of pairwise geometric relations base on＂a contrario＂formulation of RANSAC,which avoid from setting an unreliable threshold,so it can increase the accuracy of the global rotations.As to computation of the global translations,an algorithm combined golden section with linear programming is proposed to accurately compute the relative translations vector in triple of views,so it can be used to improve the accuracy of global translations.Testing results shows that our algorithm can effectively improve the accuracy of the 3 Dreconstruction.