Scholarlink

<p>Correspondence-based Point Cloud Registration (PCR) is crucial for 3D visualization applications, especially in change detection. Most PCR models depend on precise initialization using a set of closest points to establish correspondences, a method that often fails due to random variations in point positions and the influence of outliers. The presence of noise and outliers significantly compromises the quality of initial correspondences, leading to inaccuracies in alignment. While some correspondence-prediction methods inspired by nonconvex techniques show promise, they remain sensitive to the underlying data structure and are not well-suited for complex scenarios involving dynamic point clouds. In this paper, we propose a new approach: the Correspondence Evolving Assistant Network (CEANet), a point transformer-inspired mechanism designed to enhance point cloud registration. Unlike existing methods, CEANet leverages a unique conditional correspondence-fitness function that dynamically assesses and prioritizes inliers, allowing for more robust and accurate correspondence predictions through context-aware random sampling of key points. This allows CEANet to generate accurate correspondence coordinates while effectively accounting for rotation and translation in the registration process. Additionally, the model prioritizes inliers, systematically disregarding outlier points to refine transformation calculations and update point cloud alignment. The iterative process gradually removes outliers until all points fit within a defined bounding box (bbx), ensuring robust performance even in challenging environments. Specifically, registering static-to-mobile point clouds is challenging due to temporal misalignment and varying viewpoints, but the received results are notable on SABRE and Kitti dataset. Extensive experiments proved that the proposed method produced competitive results with state-of-the-art methods, achieving accurate fitness performance of +5.18E+08 and +7.33E+04 on SABRE and Kitti, respectively, and the git <a href="https://github.com/Vladimyr23/SabreProject_code">https://github.com/Vladimyr23/SabreProject_code</a> link.</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

CEANet: An end-to-end transformer based static-mobile points alignment approach

  • 期刊论文

  • 发表日期：

    2025年5月22日

查看全文
  • Vladimir Yesipov,
  • M. S. Mekala,
  • Paul Edge,
  • Eyad Elyan

摘要

Correspondence-based Point Cloud Registration (PCR) is crucial for 3D visualization applications, especially in change detection. Most PCR models depend on precise initialization using a set of closest points to establish correspondences, a method that often fails due to random variations in point positions and the influence of outliers. The presence of noise and outliers significantly compromises the quality of initial correspondences, leading to inaccuracies in alignment. While some correspondence-prediction methods inspired by nonconvex techniques show promise, they remain sensitive to the underlying data structure and are not well-suited for complex scenarios involving dynamic point clouds. In this paper, we propose a new approach: the Correspondence Evolving Assistant Network (CEANet), a point transformer-inspired mechanism designed to enhance point cloud registration. Unlike existing methods, CEANet leverages a unique conditional correspondence-fitness function that dynamically assesses and prioritizes inliers, allowing for more robust and accurate correspondence predictions through context-aware random sampling of key points. This allows CEANet to generate accurate correspondence coordinates while effectively accounting for rotation and translation in the registration process. Additionally, the model prioritizes inliers, systematically disregarding outlier points to refine transformation calculations and update point cloud alignment. The iterative process gradually removes outliers until all points fit within a defined bounding box (bbx), ensuring robust performance even in challenging environments. Specifically, registering static-to-mobile point clouds is challenging due to temporal misalignment and varying viewpoints, but the received results are notable on SABRE and Kitti dataset. Extensive experiments proved that the proposed method produced competitive results with state-of-the-art methods, achieving accurate fitness performance of +5.18E+08 and +7.33E+04 on SABRE and Kitti, respectively, and the git https://github.com/Vladimyr23/SabreProject_code link.