Yu Chanki, Ju Da Young
Department of Media Technology, Sogang University, Seoul 04107, Korea.
Yonsei Institute of Convergence Technology, Yonsei University, Incheon 21993, Korea.
Sensors (Basel). 2018 Feb 10;18(2):544. doi: 10.3390/s18020544.
In this paper, a globally optimal algorithm based on a maximum feasible subsystem framework is proposed for robust pairwise registration of point cloud data. Registration is formulated as a branch-and-bound problem with mixed-integer linear programming. Among the putative matches of three-dimensional (3D) features between two sets of range data, the proposed algorithm finds the maximum number of geometrically correct correspondences in the presence of incorrect matches, and it estimates the transformation parameters in a globally optimal manner. The optimization requires no initialization of transformation parameters. Experimental results demonstrated that the presented algorithm was more accurate and reliable than state-of-the-art registration methods and showed robustness against severe outliers/mismatches. This global optimization technique was highly effective, even when the geometric overlap between the datasets was very small.
本文提出了一种基于最大可行子系统框架的全局最优算法，用于点云数据的鲁棒成对配准。配准被公式化为一个混合整数线性规划的分支定界问题。在所提出的算法中，在两组距离数据之间的三维（3D）特征的假定匹配中，在存在错误匹配的情况下找到最大数量的几何正确对应关系，并以全局最优方式估计变换参数。该优化不需要变换参数的初始化。实验结果表明，所提出的算法比现有配准方法更准确、可靠，并且对严重的离群值/不匹配具有鲁棒性。即使数据集之间的几何重叠非常小，这种全局优化技术也非常有效。
