Scholarlink

<p>When dealing with low-quality point clouds or low overlap areas, there are often issues of limited accuracy and difficult registration. The discriminative ability of local descriptors is significantly affected by the quality of point clouds, especially sensitive to issues such as partial occlusion and uneven density. This article proposes a globally robust registration algorithm for occlusion and low overlap point clouds to address this issue. Firstly, by improving local descriptors based on LDASH, the robustness of descriptors to occlusion and uneven density is enhanced; At the same time, a chi square distance metric designed specifically for histogram descriptors is introduced to further improve the effectiveness of descriptor matching. Next, a consensus set filter and a matching aggregator are developed to eliminate erroneous correspondences and reconstruct accurate global correspondences, in order to address the challenge of obtaining correct matches in low overlap situations. Finally, the optimal transformation matrix is obtained through optimization calculations. The experimental results show that it can achieve point cloud registration with an overlap rate as low as 0.21. The experiments on B3R, U3M, K3R, and S3R datasets demonstrate that in the low overlap rate (0.5 to 0.21) point cloud registration problem, the rotation error can be stably controlled within 5 °. Compared to traditional corresponding propagation mechanisms, the algorithm proposed in this paper can achieve a time efficiency improvement of nearly 4 to 10 times, while achieving a search accuracy of 85% or even 100% for correct matching relationships. The link to obtain the code for this article is: <a href="https://github.com/wxy4845/Robust-Descriptor-and-Hierarchical-Filtering-Aggregation.git">https://github.com/wxy4845/Robust-Descriptor-and-Hierarchical-Filtering-Aggregation.git</a>.</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

Occlusion-robust descriptor and hierarchical filtering-aggregation for 3D registration

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • ShuQiu Tan,
  • XinYu Wang,
  • QiongMin Zhang

摘要

When dealing with low-quality point clouds or low overlap areas, there are often issues of limited accuracy and difficult registration. The discriminative ability of local descriptors is significantly affected by the quality of point clouds, especially sensitive to issues such as partial occlusion and uneven density. This article proposes a globally robust registration algorithm for occlusion and low overlap point clouds to address this issue. Firstly, by improving local descriptors based on LDASH, the robustness of descriptors to occlusion and uneven density is enhanced; At the same time, a chi square distance metric designed specifically for histogram descriptors is introduced to further improve the effectiveness of descriptor matching. Next, a consensus set filter and a matching aggregator are developed to eliminate erroneous correspondences and reconstruct accurate global correspondences, in order to address the challenge of obtaining correct matches in low overlap situations. Finally, the optimal transformation matrix is obtained through optimization calculations. The experimental results show that it can achieve point cloud registration with an overlap rate as low as 0.21. The experiments on B3R, U3M, K3R, and S3R datasets demonstrate that in the low overlap rate (0.5 to 0.21) point cloud registration problem, the rotation error can be stably controlled within 5 °. Compared to traditional corresponding propagation mechanisms, the algorithm proposed in this paper can achieve a time efficiency improvement of nearly 4 to 10 times, while achieving a search accuracy of 85% or even 100% for correct matching relationships. The link to obtain the code for this article is: https://github.com/wxy4845/Robust-Descriptor-and-Hierarchical-Filtering-Aggregation.git.