面向缺失点云配准的镜像迭代最近点算法  

Mirrored Iterative Closest Point Algorithm for Missing PointCloud Registration

作　　者：徐文菲 金莉[1] 韩旭 程浩喆 田暄 XU Wenfei;JIN Li;HAN Xu;CHENG Haozhe;TIAN Xuan(School of Software Engineering,Xi’an Jiaotong University,Xi’an 710049,China)

机构地区：[1]西安交通大学软件学院,西安710049

出　　处：《西安交通大学学报》2023年第7期201-212,220,共13页Journal of Xi'an Jiaotong University

基　　金：陕西省重点研发计划资助项目(2021GY-025)。

摘　　要：面向点云配准任务,以改善重叠度较低的点云对之间的配准效果为研究目的,提出了一种有效的缺失点云配准算法——镜像迭代最近点算法。该算法以建立源点云和目标点云之间的镜像型对应关联性为核心,具体过程为:首先建立源点云到目标点云的前向对应关系,以捕获位于重叠区域的特征点;然后建立重叠区域的后向对应关系,以获取可靠匹配对集合;最后根据可靠匹配对估计最优刚体变换矩阵。此外,通过KD树构建和特征扩展两方面进行优化,以提升算法效率。所提算法仅依赖重叠区域匹配对集合,具有良好的鲁棒性和抗干扰能力。在斯坦福数据集上的实验结果表明:对于较低重叠度的数据集,所提算法在精度和效率上均优于以往的多数算法;对于较高重叠度的数据集,所提算法使精度平均提升28.8%,使效率平均提升47.9%。实验证明所提算法能快速且可靠地配准缺失点云。Focusing on point cloud registration tasks,an effective missing point cloud registration algorithm,the mirrored iterative closest point was proposed,to improve the registration between point cloud pairs with low overlap.The core of the proposed algorithm is to establish the mirror-type corresponding correlation between the source point cloud and the target point cloud.The specific process is as follows:firstly,establish a forward correspondence between the source point cloud and the target point cloud to capture the characteristic points located in the overlapping regions.Then,establish a backward correspondence of overlapping regions to obtain a collection of reliable matching pairs.In the end,estimate the optimal rigid transformation matrix based on the reliable matching pairs.In addition,the KD tree construction and feature extension are optimized to improve the algorithm efficiency.The proposed algorithm relies only on matching pair sets in overlapping regions and has excellent robustness and anti-interference.The results of experiment on Stanford dataset show that,for datasets with low overlap,the algorithm is better than the previous algorithms in accuracy and efficiency,while for datasets with high overlap,the algorithm improves the accuracy by 28.8%and the efficiency by 47.9%on average.Experiments demonstrate that the algorithm registers missing point clouds quickly and reliably.

关 键 词：缺失点云配准 镜像迭代最近点 有限重叠度 可靠匹配对 特征扩展 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]