Scholarlink

<p>We introduce an innovative fine registration framework of point clouds, namely TrCPDs, to address challenges posed by low-overlap and noisy scenarios. Most probabilistic model-based registration methods, including Coherent Point Drift (CPD), assume complete overlap between source point cloud and target point cloud, suffering from inaccurate alignments in partial-overlap situations, which exist prevalently in real-world applications. Our method tackles the point cloud registration problem in low-overlap scenarios by introducing a trimming mechanism. This mechanism trims the correspondences based on the quantified point pair relationships and overlap rate, thereby ensuring accurate registration in low-overlap conditions. To highlight the universality of the proposed method, TrCPDs can seamlessly complement existing CPD-based methods, substantially boosting their performance in low-overlap scenarios. This integration not only preserves the robustness of probabilistic model-based methods but also leverages TrCPDs’ unique capabilities to address challenges inherent in low-overlap environments. For variants lacking closed-form solutions, we introduce a damped Newton method leveraging the Armijo condition to reduce hyperparameter sensitivity and enhance algorithmic convergence. Extensive experiments reveal our proposed framework’s noticeable robustness and stability across diverse datasets.</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

TrCPDs: Trimmed coherent point drift methods for robust fine registration of low-overlap point cloud

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Yihan Pan,
  • Zhiyong Dai,
  • Jianjun Yi,
  • Yibin Zhao,
  • Liansheng Wang

摘要

We introduce an innovative fine registration framework of point clouds, namely TrCPDs, to address challenges posed by low-overlap and noisy scenarios. Most probabilistic model-based registration methods, including Coherent Point Drift (CPD), assume complete overlap between source point cloud and target point cloud, suffering from inaccurate alignments in partial-overlap situations, which exist prevalently in real-world applications. Our method tackles the point cloud registration problem in low-overlap scenarios by introducing a trimming mechanism. This mechanism trims the correspondences based on the quantified point pair relationships and overlap rate, thereby ensuring accurate registration in low-overlap conditions. To highlight the universality of the proposed method, TrCPDs can seamlessly complement existing CPD-based methods, substantially boosting their performance in low-overlap scenarios. This integration not only preserves the robustness of probabilistic model-based methods but also leverages TrCPDs’ unique capabilities to address challenges inherent in low-overlap environments. For variants lacking closed-form solutions, we introduce a damped Newton method leveraging the Armijo condition to reduce hyperparameter sensitivity and enhance algorithmic convergence. Extensive experiments reveal our proposed framework’s noticeable robustness and stability across diverse datasets.