改进拟牛顿算法的点云稠密化应用研究

Research on application of improved Quasi-Newton algorithm in point cloud densification

针对在三维重建中,利用图像序列重建的方法耗时过长、模型精度低的问题,该文提出一种对离散点云稠密化的优化方法。以图像序列的三维模型重建为背景,在增量式运动恢复结构算法(SfM)生成三维稀疏点云的基础上,对目标物体表面的稀疏离散点进行稠密化,并提出了一种针对基于面片的多视角立体几何算法(PMVS)时间复杂度的优化方法。针对稠密点云重建过程中PMVS算法耗时过长的问题,构建一种基于原始的拟牛顿优化方法的改进算法。在PMVS算法的面片优化部分中,对拟牛顿算法(BFGS)的修正矩阵迭代公式进行改进,确保全局收敛的同时也提高了收敛速度。实验结果表明,该文使用的改进算法不仅加快了三维重建的速度,而且适当提升了稠密点云重建的质量。

Aiming at the problem that the method of using image sequence reconstruction in 3 D reconstruction is too long and the model accuracy is low,an optimization method for point cloud densification in 3 D reconstruction is proposed.Based on the reconstruction of the 3 D model of the image sequence,this paper uses the incremental motion restoration structure algorithm(SfM)to generate 3 D sparse point clouds,densifies the sparse discrete points on the surface of the target object,and proposes a method for PMVS The optimization method of algorithm time complexity.Aiming at the problem that the surface-based multi-view stereo geometry algorithm(PMVS)takes too long in the process of dense point cloud reconstruction,an improved algorithm based on the original quasi-Newton optimization method is constructed.In the patch optimization part of the PMVS algorithm,the modified matrix iteration formula of the Quasi-Newton optimization algorithm is improved to ensure global convergence while increasing the convergence speed.The experimental results show that improved Quasi-Newton algorithm used in this paper not only accelerates the speed of 3 D reconstruction,but also appropriately improves the quality of dense point cloud reconstruction.