《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (8): 2584-2593.DOI: 10.11772/j.issn.1001-9081.2025070887

• 多媒体计算与计算机仿真 • 上一篇    下一篇

双分支结构下多层次特征融合的点云配准网络

刘明1,2, 沈东奇1,2(), 孟子洋1,2   

  1. 1.云南民族大学 电气信息工程学院,昆明 650504
    2.云南省无人自主系统重点实验室(云南民族大学),昆明 650504
  • 收稿日期:2025-08-04 修回日期:2025-09-29 接受日期:2025-10-10 发布日期:2025-11-05 出版日期:2026-08-10
  • 通讯作者: 沈东奇
  • 作者简介:刘明(1973—),男,云南曲靖人,研究员,博士,主要研究方向:自动控制与导航、机器人SLAM算法
    沈东奇(1998—),男,山东淄博人,硕士研究生,CCF会员,主要研究方向:激光SLAM算法、点云配准算法
    孟子洋(2000—),男,山东青岛人,硕士研究生,主要研究方向:SLAM算法。
  • 基金资助:
    国家自然科学基金资助项目(52061042)

Point cloud registration network with dual-branch multi-level feature fusion

Ming LIU1,2, Dongqi SHEN1,2(), Ziyang MENG1,2   

  1. 1.School of Electrical and Information Engineering,Yunnan Minzu University,Kunming Yunnan 650504,China
    2.Yunnan Province Key Laboratory of Unmanned Autonomous Systems (Yunnan Minzu University),Kunming Yunnan 650504,China
  • Received:2025-08-04 Revised:2025-09-29 Accepted:2025-10-10 Online:2025-11-05 Published:2026-08-10
  • Contact: Dongqi SHEN
  • About author:LIU Ming, born in 1973, Ph. D., research fellow. His research interests include automatic control and navigation, robot SLAM algorithms.
    MENG Ziyang, born in 2000, M. S. candidate. His research interests include SLAM algorithms.
  • Supported by:
    National Natural Science Foundation of China(52061042)

摘要:

针对点云配准中因部分重叠、遮挡与噪声干扰导致的配准精度差与鲁棒性不足的问题,提出一种双分支结构下的多层次特征融合的点云配准网络DMFNet(Dual-branch Multi-level Feature Fusion Network)。该网络在编码阶段并行设置旋转分支与平移分支,并在浅层、中层与深层分别引入自注意力融合与交叉注意力融合模块,从而实现源点云与参考点云之间的多尺度特征交互与深度融合,同时,设计了旋转与平移特征融合模块以增强姿态估计能力;在回归阶段基于轻量级Set Transformer回归器,采用多层诱导注意力块与注意力池化模块来直接回归四元数与平移向量。DMFNet不依赖重叠区域的检测或显式掩码估计,具备更强的适应性与泛化能力。在ModelNet40数据集上与6种点云配准方法进行对比实验,并在斯坦福3D扫描数据集进行泛化能力实验。实验结果表明,在ModelNet40数据集上添加噪声的情况下二次采样,相较于最大极大团配准(MAC)方法,DMFNet在RMSE(t)和Error(R)指标上分别降低了21.32%和14.47%,展现出更优的鲁棒性与配准精度。

关键词: 深度学习, 点云配准, 三维点云, 特征融合, 注意力机制

Abstract:

A point cloud registration network with dual-branch multi-level feature fusion, DMFNet (Dual-branch Multi-level Feature Fusion Network), was proposed to address poor registration accuracy and insufficient robustness problems in point cloud registration caused by partial overlapping, occlusion, and noise interference. In the network, during the encoding stage, a rotation branch and a translation branch were configured in parallel, and self-attention fusion as well as cross-attention fusion modules were inserted at shallow, middle, and deep layers, so as to achieve multi-scale feature interaction and deep fusion between source and reference point clouds. At the same time, a rotation-translation feature fusion module was designed to strengthen pose estimation capability. During the regression stage, based on the lightweight Set Transformer regressor, multi-layer induced-attention blocks and an attention-pooling module were used to regress the quaternion and the translation vector directly. DMFNet does not rely on overlap-region detection or explicit mask estimation, and it shows strong adaptability and generalization. Comparative experiments with six point cloud registration methods were conducted on the ModelNet40 dataset, and a generalization study was carried out on the Stanford 3D scanning dataset. The results show that on the ModelNet40 dataset under noisy conditions and two-time sampling, DMFNet reduces RMSE(t) and Error(R) by 21.32% and 14.47% compared with MAximum Clique-based registration (MAC) method, respectively, and demonstrates superior robustness and registration accuracy.

Key words: deep learning, point cloud registration, 3D point cloud, feature fusion, attention mechanism

中图分类号: