《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (11): 3573-3582.DOI: 10.11772/j.issn.1001-9081.2024111625

• 人工智能 • 上一篇    

面向鲁棒点云配准的邻域关注和拓扑感知的图卷积方法

徐梦楠1, 叶海良1, 曹飞龙2()   

  1. 1.中国计量大学 理学院,杭州 310018
    2.浙江师范大学 数学与交叉科学研究院,杭州 310012
  • 收稿日期:2024-11-18 修回日期:2025-03-25 接受日期:2025-04-16 发布日期:2025-04-22 出版日期:2025-11-10
  • 通讯作者: 曹飞龙
  • 作者简介:徐梦楠(1997—),女,浙江杭州人,硕士,主要研究方向:深度学习、图神经网络、点云配准
    叶海良(1990—),男,浙江绍兴人,副教授,博士,CCF会员,主要研究方向:深度学习、点云分析
  • 基金资助:
    国家自然科学基金资助项目(62032022);国家自然科学基金资助项目(62176244)

Neighborhood-attention and topology-aware graph convolution method for robust point cloud registration

Mengnan XU1, Hailiang YE1, Feilong CAO2()   

  1. 1.College of Sciences,China Jiliang University,Hangzhou Zhejiang 310018,China
    2.Institute of Mathematics and Cross-disciplinary Science,Zhejiang Normal University,Hangzhou Zhejiang 310012,China
  • Received:2024-11-18 Revised:2025-03-25 Accepted:2025-04-16 Online:2025-04-22 Published:2025-11-10
  • Contact: Feilong CAO
  • About author:XU Mengnan, born in 1997, M. S. Her research interests include deep learning, graph neural network, point cloud registration.
    YE Hailiang, born in 1990, Ph. D., associate professor. His research interests include deep learning, point cloud analysis.
  • Supported by:
    National Natural Science Foundation of China(62032022)

摘要:

现有的大部分点云配准方法通常忽略了邻域内相邻节点之间的关系,导致对局部几何结构特征提取不足。针对该问题,提出一种面向鲁棒点云配准的邻域关注和拓扑感知的图卷积(NATA)方法,以捕捉更深层次的语义特征和更丰富的几何信息。首先,设计了级联几何感知模块,该模块利用基于自注意力的局部邻域更新图卷积模块,关注局部图的内在几何结构,以获得更精确的局部拓扑信息;其次,级联结构组合不同维度的局部拓扑信息,以产生更具判别性的局部描述符;最后,提出特征交互图更新模块,该模块在点云中建立了一种注意力机制来捕捉点云的隐含关系并感知点云的形状特征。在具有挑战性的3D点云基准测试上的实验结果表明,所提方法在部分噪声点云配准中的平均绝对误差(MAE)在未知形状和未知类别下分别取得了0.157 2和0.154 4的优异结果。

关键词: 深度学习, 注意力机制, 点云, 配准, 图卷积网络

Abstract:

Most existing point cloud registration methods often ignore the relationship between neighboring nodes in the neighborhood, resulting in insufficient feature extraction of local geometric structures. To solve this problem, a Neighborhood-Attention and Topology-Aware graph convolution (NATA) method for robust point cloud registration was proposed to capture deeper semantic features and richer geometric information. Firstly, a cascade geometry-aware module was designed, which used a self-attention-based local neighborhood update graph convolution module to focus on the intrinsic geometric structure of local graphs, thereby obtaining more accurate local topological information. Secondly, a cascade structure combined various levels of local topological information to produce a more discriminative collection of local descriptors. Finally, a feature interaction-graph update module was proposed, which created an attention mechanism in the point clouds to capture their implicit relationships and perceive shape features of the point clouds. Experimental results on a challenging 3D point cloud benchmark test show that the proposed method achieves excellent Mean Absolute Error (MAE) of 0.157 2 and 0.154 4 for partial noisy point clouds registration under unknown shapes and unknown categories, respectively.

Key words: deep learning, attention mechanism, point cloud, registration, Graph Convolutional Network (GCN)

中图分类号: