<p>Multi-view human point cloud registration is a crucial step in 3D human reconstruction tasks. The symmetric structures and similar geometric features in human point clouds often lead to feature mismatches in point cloud registration. Therefore, we propose a pipeline for game tree registration based on semantic constraints and feature weighting (GTR-SCFW) that enhances the stability and accuracy of feature matching, thereby improving the registration precision of multi-view point clouds. First, we calculate and compare the feature similarity between multi-view point clouds and use a generalized best-first search (BFS) method to construct a multi-layered registration game tree. At each game node, overlapping regions are divided into multiple sub-regions based on semantic information, and global fast registration is used to determine the matching relationships of features within each sub-region. Then, the best matching points in each sub-region are selected based on the confidence of feature pairs, and the weights of all the best point pairs are calculated. Finally, the initial rigid transformation matrix is computed using weighted least squares (WLS), and ICP is employed to achieve fast fine registration. GTR-SCFW effectively avoids incorrect matching relationships caused by geometric feature similarity during the initial transformation estimation, providing a good initial pose for iterative closest point (ICP) fine registration. For point clouds with different initial poses, the registration’s rotational error approaches 0<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10489_2025_6296_Article_IEq1.gif" Format="GIF" Height="7" Rendition="HTML" Resolution="72" Type="Linedraw" Width="9" /> </InlineMediaObject> <EquationSource Format="TEX">\(^\circ \)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mo>∘</mo> </mmultiscripts> </math></EquationSource> </InlineEquation>, while the translational error is as low as 1.203e-4&#xa0;mm. Comparative experimental results show that this method outperforms existing feature-based registration methods regarding robustness, reliability, and computational efficiency.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Multi-view human point cloud registration method with overlapping regions semantic constraints and feature weighting

  • Ming Li,
  • Guiqin Li,
  • Xihang Li,
  • Tiancai Li

摘要

Multi-view human point cloud registration is a crucial step in 3D human reconstruction tasks. The symmetric structures and similar geometric features in human point clouds often lead to feature mismatches in point cloud registration. Therefore, we propose a pipeline for game tree registration based on semantic constraints and feature weighting (GTR-SCFW) that enhances the stability and accuracy of feature matching, thereby improving the registration precision of multi-view point clouds. First, we calculate and compare the feature similarity between multi-view point clouds and use a generalized best-first search (BFS) method to construct a multi-layered registration game tree. At each game node, overlapping regions are divided into multiple sub-regions based on semantic information, and global fast registration is used to determine the matching relationships of features within each sub-region. Then, the best matching points in each sub-region are selected based on the confidence of feature pairs, and the weights of all the best point pairs are calculated. Finally, the initial rigid transformation matrix is computed using weighted least squares (WLS), and ICP is employed to achieve fast fine registration. GTR-SCFW effectively avoids incorrect matching relationships caused by geometric feature similarity during the initial transformation estimation, providing a good initial pose for iterative closest point (ICP) fine registration. For point clouds with different initial poses, the registration’s rotational error approaches 0 \(^\circ \) , while the translational error is as low as 1.203e-4 mm. Comparative experimental results show that this method outperforms existing feature-based registration methods regarding robustness, reliability, and computational efficiency.