实现激光点云高效配准的ICP优化及性能验证

激光点云常规匹配算法是迭代最近点(Iterative Closest Point, ICP)算法，但其收敛速度慢、鲁棒性差，因此，提出一种融合多种优化算法的激光点云高效ICP配准方法。首先对点云体素滤波降采样，通过ISS算子提取关键点，采用快速点特征直方图(Fast Point Feature Histograms, FPFH)提取关键点特征，嵌入多核多线程并行处理模式 (OpenMP)提高特征提取速度；然后基于提取的FPFH特征，使用采样一致性初始配准算法(Sample Consensus Initial Alignment, SAC-IA)进行相似特征点粗配准，获取点云集间的初始旋转平移变换矩阵；最后采用ICP算法进行精配准，同时采用最优节点优先(Best Bin First, BBF)优化K-D tree近邻搜索法来加速对应关系点对的搜索，并设定动态阈值消除错误对应点对，提高配准快速性和准确性。对两个实例的配准点云进行了实验验证，结果表明，提出的优化配准算法具有明显速度优势和精度优势。

Optimization and performance verification of high efficiency ICP registration for laser point clouds

The conventional Iterative Closest Point(ICP) matching algorithm for laser point clouds had problems of slow convergence and poor robustness, therefore, a point clouds registration method combining multiple optimization methods was proposed. Firstly, point clouds were de-sampled using voxel grid filtering and key points were extracted by ISS operator, then feature extraction algorithm was performed to obtain Fast Point Feature Histograms(FPFH) features of key points, and the multi-core and multi-thread OpenMP parallel processing mode was operated to improve the speed of feature extraction. Then, based on the extracted FPFH features, the Sample Consistency Initial Alignment(SAC-IA) algorithm was used for coarse registration of similar feature points to obtain initial transformation matrix between point clouds sets. Finally, the ICP algorithm was used for fine registration, and the K-D tree nearest neighbor search method optimized by Best Bin First(BBF) was used to accelerate the search speed of corresponding point pairs, and dynamic threshold was set to eliminate the wrong corresponding point pairs, so as to improve the speed and accuracy of point clouds registration. Experimental research on two sets of point clouds shows that the optimized registration algorithm has obvious speed advantages and improves the registration accuracy.