基于改进RPM-Net的船舶分段点云模型配准方法  

Ships Registration Method of Ship Segmented Point Cloud Model Based on Improved RPM-Net

作　　者：刘欣阳 苌道方[1] 马彦军 杜吉旺 LIU Xinyang;CHANG Daofang;MA Yanjun;DU Jiwang(Logistics Engineering College,Shanghai Maritime University,Shanghai 201306,China;Hudong-Zhonghua Shipbuiding(Group)Co.,Ltd.,Shanghai 200129,China)

机构地区：[1]上海海事大学物流工程学院,上海201306 [2]沪东中华造船(集团)有限公司,上海200129

出　　处：《船舶工程》2025年第1期132-139,共8页Ship Engineering

摘　　要：[目的]船舶分段数据具有内部构造复杂、粗差点较多、位姿随机等特点,现有算法的配准精度较差,为解决这一问题,[方法]提出一种基于改进RPM-Net的使用快速点特征直方图的鲁棒点匹配(RPM-FPFH)算法。在特征提取阶段引入快速点特征直方图(FPFH)算法以充分捕捉分段内部复杂的几何特征。将分段特征混合并输入Transformer模块,共同学习局部与全局特征,以应对粗差数据影响。在匹配矩阵计算模块中引入熵正则化,减少初始位姿随机性对参数迭代的影响。[结果]结果表明:所提算法在船舶分段配准结果的相对旋转误差、相对平移误差与倒角距离方面的误差相比于现有算法中最优者分别减少3.89%、9.31%和2.69%。[结论]研究成果可为船舶分段配准提供一定参考。[Purpose]Ship segmentation data has the characteristics of complex internal structure,many rough spots,and random pose.The registration accuracy of existing algorithms is poor.To solve this problem,[Method]Robust Point Matching using Fast Point Feature Histograms(RPM-FPFH)is proposed based on the improved RPM Net.In the feature extraction stage,the Fast Point Feature Histograms(FPFH)algorithm is introduced to fully capture the complex geometric features within the segments.The segmented features are mixed and input into the Transformer module to jointly learn local and global features to cope with the effects of coarse data.Entropy regularization is introduced into the matching matrix calculation module to reduce the impact of initial pose randomness on parameter iteration.[Result]The results show that the proposed algorithm reduces the relative rotation error,relative translation eror,and chamfer distance error in ship segmentation registration results by 3.89%,9.31%and 2.69%,respectively,compared to the best existing algorithms.[Conclusion]The research results can provide some references for ship segmentation registration.

关 键 词：数字化造船 深度学习 船舶分段 三维点云 精度控制 

分 类 号：U671.99[交通运输工程—船舶及航道工程]