Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (12): 252-260.DOI: 10.3778/j.issn.1002-8331.2304-0089

• Graphics and Image Processing • Previous Articles     Next Articles

Research on Partial Overlapping Point Cloud Registration Algorithm for Matching Geometric Features

HU Jianghao, WANG Feng   

  1. School of Information Engineering, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2024-06-15 Published:2024-06-14

匹配几何特征的部分重叠点云配准算法研究

胡江豪,王丰   

  1. 广东工业大学 信息工程学院,广州 510006

Abstract: In order to solve the problems of outlier, redundant points and fuzzy features in the registration of partially overlapping point clouds, a new registration algorithm for partially overlapping point clouds matching geometric features is proposed. Feature interaction and multilayer perceptron are used to calculate the overlap score and feature salient value of each point to be registered, and extract the salient feature points in the overlapping area. Geometric features are captured according to length and angle, representative feature descriptors are extracted, special geometric feature matching networks are designed, internal values and outlier of key points are identified, and outlier are filtered. The registration results are obtained using weighted singular value decomposition operations. Experimental results show that for ModelNet40 dataset, compared with the benchmark algorithm, the root mean square error and mean absolute error of the proposed algorithm in rotation and translation are reduced by 59%, 45%, 83% and 66% respectively. For the ShapeNetCore dataset, the algorithm is reduced by 63%, 32%, 78%, and 50% in four indicators, respectively.

Key words: machine vision, point cloud registration, overlapping area, feature matching

摘要: 为解决部分重叠点云配准任务中异常值、冗余点和模糊特征等问题,提出了一种匹配几何特征的新型部分重叠点云配准算法。使用特征交互及多层感知器,计算待配准各点的重叠分数和特征显著值,提取重叠区域的显著特征点。根据长度和角度来捕捉几何特征,提取具有代表性的特征描述符,设计专用的几何特征匹配网络,标识关键点的内部值和异常值,过滤异常值。利用加权奇异值分解运算得到配准结果。实验结果表明,针对ModelNet40数据集,相比于基准算法,所提算法在旋转和平移上的均方根误差和平均绝对误差分别减少了59%、45%、83%、66%;针对ShapeNetCore数据集,该算法在四项指标上分别减少了63%、32%、78%、50%。

关键词: 机器视觉, 点云配准, 重叠区域, 特征匹配