首页|基于空间可变形Transformer的三维点云配准方法

基于空间可变形Transformer的三维点云配准方法

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
针对低重叠场景下点云配准方法鲁棒性差、配准精度低的问题，提出一种基于空间可变形Transformer(SDT)的三维点云配准方法。设计多级分辨率特征的提取与融合方法，显式计算点云的局部空间关系。利用SDT模块增强点云空间特征的表达能力，聚合局部与全局的特征得到特征矩阵。计算两个特征矩阵的相似度矩阵并额外地为其添加边缘松弛块，有效降低了不可行匹配对配准鲁棒性的影响，同时对相似度矩阵进行归一化等计算得到软对应置信度矩阵，根据预测的对应点空间特征是否一致来寻找点云在低重叠场景下更精确的对应关系，使用直接定义在对应关系上的损失来训练网络，将软对应关系转换为一对一的硬匹配关系，最终通过随机抽样一致性刚性变换求解器执行配准。实验结果表明，在重叠率低于30%的3DLoMatch场景中，该方法的特征匹配召回率和配准召回率相比于高度关注重叠区域的成对点云配准等方法至少提高了3。7和3。9个百分点，并且具有较强的鲁棒性。
Three-Dimensional Point Cloud Registration Method Based on Spatial Deformable Transformer
Aiming at the problem of poor robustness and low registration accuracy of point cloud registration algorithms in low overlap scenarios,this study proposes a Three-Dimensional(3D)point cloud registration method based on a Spatial Deformable Transformer(SDT).A multi-level resolution feature extraction and fusion method is designed to explicitly compute the local spatial relationships of point clouds.The SDT module is used to enhance the expressive power of the spatial features of the point cloud,and local and global features are aggregated to obtain the feature matrix.This method computes the similarity matrix of the two feature matrices and adds an edge slack block to it,effectively reducing the impact of infeasible matching on the robustness of registration.Additionally,the similarity matrix is normalized and calculated to obtain the soft correspondence confidence matrix,and the more accurate correspondence of the point cloud in the low overlap scenario is determined according to whether the spatial features of the corresponding points are consistent.The loss defined directly on the correspondence is used to train the network to convert the soft correspondence into a one-to-one hard matching relation,and finally,the registration is performed by the RANdom SAmple Consensus(RANSAC)rigid transformation solver.Experimental results demonstrate that the Feature Matching Recall(FMR)and Registration Recall(RR)of the proposed method are at least 3.7 and 3.9 percentage points higher than those of Pairwise pointcloud REgistration with Deep Attention To the Overlap Region(PREDATOR),respectively,including other methods in the 3DLoMatch scenario with less than 30%overlap,with strong robustness.

low overlap ratemulti-feature fusiondeformable self-attentionedge slack blockoverlap correspondence prediction

谢帅康、熊风光、朱新杰、宋宁栋、李文清、王廷凤

展开 >

中北大学计算机科学与技术学院,山西 太原 030051

中北大学山西省视觉信息处理及智能机器人工程研究中心,山西 太原 030051

中北大学机器视觉与虚拟现实山西省重点实验室,山西 太原 030051

低重叠率 多特征融合 可变形自注意力 边缘松弛块 重叠对应预测

国家自然科学基金山西省回国留学人员科研基金山西省科技成果转化引导专项基金山西省科技重大专项计划"揭榜挂帅"项目

622724262020-113202104021301055202201150401021

2024

10.19678/j.issn.1000-3428.0067566

计算机工程
华东计算技术研究所　上海市计算机学会

计算机工程

CSTPCD北大核心
影响因子：0.581
ISSN：1000-3428
年,卷(期)：2024.50(3)
  • 31