Scholarlink

<p>Point cloud registration plays an important role in 3D computer vision such as robotics and 3D reconstruction. The iterative closest point (ICP)-based registration method may lead to falling into the local optimum, slow convergence speed, and mismatch for point clouds with low-overlap. In this paper, a new robust and fast point cloud registration method named MSKICP is proposed to effectively solve this problem. Firstly, a multiscale descriptor that can quickly extract the local structure of uniform sampling points is computed, and the nearest neighbor search is adopted to form seed match based on curvature feature. Secondly, the overlapping region and correspondence are acquired by clustering and chunking (CAC) strategy. Finally, a kernel mean <i>p</i>-power error (KMPE) function-based error metric function is established, which suppresses outliers by reasonably allocating weights. Furthermore, the KMPE error metric is extended to point-to-plane ICP to achieve registration. Experimental results show that MSKICP can achieve high registration accuracy for point cloud with low-overlap. More encouragingly, it is rather robust to outliers.</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

MSKICP: multiscale descriptor and KMPE kernel function-improved iterative closest point for point cloud registration

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Shengmei Chen,
  • Hao Deng,
  • Jingyi Han,
  • Cheng Liu,
  • Jinye Peng,
  • Lin Wang

摘要

Point cloud registration plays an important role in 3D computer vision such as robotics and 3D reconstruction. The iterative closest point (ICP)-based registration method may lead to falling into the local optimum, slow convergence speed, and mismatch for point clouds with low-overlap. In this paper, a new robust and fast point cloud registration method named MSKICP is proposed to effectively solve this problem. Firstly, a multiscale descriptor that can quickly extract the local structure of uniform sampling points is computed, and the nearest neighbor search is adopted to form seed match based on curvature feature. Secondly, the overlapping region and correspondence are acquired by clustering and chunking (CAC) strategy. Finally, a kernel mean p-power error (KMPE) function-based error metric function is established, which suppresses outliers by reasonably allocating weights. Furthermore, the KMPE error metric is extended to point-to-plane ICP to achieve registration. Experimental results show that MSKICP can achieve high registration accuracy for point cloud with low-overlap. More encouragingly, it is rather robust to outliers.