基于分层穿插多视角点云配准算法研究
Research on Point Cloud Registration Algorithm Based on Layered Interpolation and Multiple Perspectives
周文亚 1马行 2穆春阳 3胡冲1
作者信息
- 1. 北方民族大学电气信息工程学院,银川 750021
- 2. 北方民族大学电气信息工程学院,银川 750021;北方民族大学宁夏智能信息与大数据处理重点实验室,银川 750021
- 3. 北方民族大学宁夏智能信息与大数据处理重点实验室,银川 750021;北方民族大学机电工程学院,银川 750021
- 折叠
摘要
为解决毛胚铸件三维重建因多视角配准过程中引起的累积误差大、配准精度差以及配准效率低的问题,提出了一种分层穿插多视角点云配准算法.首先,利用工业相机对铸件物体获取不同视角下的点云信息进行预处理;其次,通过邻域法向量夹角特性提取特征点,并对其进行FPFH特征点描述,对每层相邻点云用改进的4PCS算法进行粗配准,获得良好的初始位置,再利用KD-tree加速搜索的广义迭代最近点GICP算法进行精配准;最后,利用分层穿插多视角配准策略拼接成完整的点云.实验结果表明,提出的配准方法相对于SAC-IA+ICP算法、4PCS+ICP算法、Super-4PCS+ICP算法,配准时间分别降低了96.5%、96.1%、88.3%,配准精度分别提高了79.5%、71.5%、55.1%.此算法配准质量不仅优于顺序多视角配准的质量,而且在鲁棒性和精确配准方面有明显优势,对后续三维重建提供了一种高效的方法.
Abstract
To solve the problems of large cumulative errors,poor registration accuracy,and low registration efficiency caused by multi view registration in 3D reconstruction of rough castings,this paper proposes a lay-ered interpolation multi view point cloud registration algorithm.Firstly,industrial cameras are used to prepro-cess point cloud information from different perspectives on casting objects;Secondly,feature points are ex-tracted through the angle characteristics of neighborhood normal vectors,and FPFH feature point descriptions are performed on them.An improved 4PCS algorithm is used for coarse registration of adjacent point clouds in each layer to obtain good initial positions.Then,the generalized iterative nearest point GICP algorithm ac-celerated by KD tree is used for fine registration;Finally,a layered interpolation multi view registration strate-gy is used to stitch together a complete point cloud.The experimental results show that the registration meth-od proposed in this paper reduces registration time by 96.5%,96.1%,and 88.3%respectively compared to the SAC-IA+ICP algorithm,4PCS+ICP algorithm,and Super-4PCS+ICP algorithm,and improves reg-istration accuracy by 79.5%,71.5%,and 55.1%,respectively.The registration quality of the algorithm in this article is not only better than that of sequential multi view registration,but also has obvious advantages in robustness and accurate registration,providing an efficient method for subsequent 3D reconstruction.
关键词
三维点云配准/多视角拼接/邻域特征点提取/4PCSKey words
3D point cloud registration/multi perspective stitching/neighborhood feature point extraction/4PCS引用本文复制引用
基金项目
自治区科技创新领军人才培养工程项目(2021GKLRLX08)
银川市科技创新项目(2022GX04)
北方民族大学研究生创新项目(YCX24106)
出版年
2024
NETL
NSTL
万方数据