一种航空影像的三维重建全局优化方法

A global optimization method for 3D reconstruction of aerial images

针对大规模航空影像三维重建全局优化的存储开销大、无法整体求解的问题,提出了一种改进的全局优化方法。该方法基于同名点对应的同名光线与基线共面这一约束,以相机参数为未知量,由影像间的同名点建立约束方程组,运用非线性最小二乘法迭代求解,最终得到各相机优化后的参数。与传统光束法平差相比,在约束方程中消除了空间点三维坐标,大幅减少了未知量,降低了优化时的存储开销和计算量;同时在构造约束方程组时,直接生成法矩阵,使矩阵规模与同名点数量无关,进一步降低了存储开销,从而能够联立大量影像、引入更多的同名点进行全局优化。基于序列航空影像、采用本方法和SBA源码包进行对比实验,结果表明,在近似垂直摄影且影像规模在千幅级情况下,本方法的内存开销较SBA显著降低,单次迭代时间和收敛速度均优于SBA,且相对精度和绝对精度亦优于SBA,证明利用两视图共面约束进行三维重建全局优化和摄影测量加密的可行性、有效性和优势。

In view of the defect of too large storage cost to be solved as a whole in the optimization process of aerial imagery-based 3D reconstruction,a global optimization approach applying coplanar constraint was proposed.The constraint equations,in which the camera parameters were the unknows,were constructed from the corresponding points according to the coplanar constraint among the two corresponding optical rays and the base-line.The constraint equations were solved iteratively through non-linear least square method,and the refined camera parameters were yielded finally.Compared with the traditional bundle adjustment,the 3D coordinates of space points were eliminated in the constraint equation,which greatly decreased the unknowns and subsequently reduced the storage cost and calculation amount during optimization.At the same time,the normal matrix was directly generated when constructing the constraint equations,making the matrix size independent of the number of corresponding points,further reducing the storage cost,so that a large number of images could be simultaneously used,and more corresponding points could be introduced in the global optimization.Experiments were conducted between the proposed approach and SBA source code package based on sequential aerial images.Experimental results showed that under the condition of approximate vertical photography and image scale of 1000 frames,the memory cost of the proposed approach was significantly reduced compared with SBA,the single iteration time and convergence speed were better than SBA,and the relative and absolute accuracy were also better than SBA,which proved the feasibility,effectiveness and advantage of global optimization applying two-view coplanar constraints in 3D reconstruction and photogrammetric point determination.