测绘学报 ›› 2026, Vol. 55 ›› Issue (10): 1829-1838.doi: 10.11947/j.AGCS.2026.20260187

• 摄影测量学与遥感 • 上一篇    下一篇

面向大规模激光雷达点云语义分割的长序列几何增强网络

黄鹤(), 张鑫泰, 张译丹, 杨军星()   

  1. 北京建筑大学测绘与城市空间信息学院,北京 102616
  • 收稿日期:2026-05-13 修回日期:2026-06-16 出版日期:2026-11-16 发布日期:2026-11-16
  • 通讯作者: 杨军星 E-mail:huanghe@bucea.edu.cn;yangjunxing@bucea.edu.cn
  • 作者简介:黄鹤(1977—),男,博士,教授,研究方向为导航定位。E-mail:huanghe@bucea.edu.cn
  • 基金资助:
    国家自然科学基金(42201483; 42501568)

Long-sequence geometry enhancement network for large-scale LiDAR point cloud semantic segmentation

He Huang(), Xintai Zhang, Yidan Zhang, Junxing Yang()   

  1. School of Geomatics and Urban Spatial Informations, Beijing University of Civil Engineering and Architecture, Beijing 102616 China
  • Received:2026-05-13 Revised:2026-06-16 Online:2026-11-16 Published:2026-11-16
  • Contact: Junxing Yang E-mail:huanghe@bucea.edu.cn;yangjunxing@bucea.edu.cn
  • About author:Huang He (1977—), male, PhD, professor, majors in navigation and positioning. E-mail: huanghe@bucea.edu.cn
  • Supported by:
    The National Natural Science Foundation of China(42201483; 42501568)

摘要:

大规模激光雷达点云语义分割是城市三维场景建模、道路地物提取和自动驾驶环境感知的重要基础。现有基于选择性状态空间模型的Mamba点云方法能够以线性复杂度建模长程依赖,但点云序列化易削弱原始邻域关系,导致道路标线、围栏、杆状物和地物边界等细长结构表达不足。针对该问题,本文提出一种融合局部几何增强的Mamba点云语义分割网络GradMamba-Net。该网络分序列化前、序列编码和解码恢复3个阶段协同设计:首先,利用局部梯度嵌入模块(LGEM)增强中心点与邻域点之间的差分几何表达;然后,在Mamba编码过程中引入多尺度空间聚合模块,补充一维序列建模中易弱化的局部空间结构;最后,通过跨通道混洗重构解码器加强深层语义特征与浅层几何细节之间的信息交互。试验在Toronto-3D和WHU-Urban3D数据集上开展,其中在Toronto-3D数据集上取得82.3%的mIoU,较3D-UMamba提升4.1个百分点;在WHU-Urban3D的MLS-S和ALS-S场景上分别取得61.2%和64.5%的mIoU。结果表明,本文方法能够提升细长地物和边界区域的分割效果,并保持较高的推理效率。

关键词: LiDAR点云, 语义分割, Mamba, 长序列建模, 局部几何增强, 特征重构

Abstract:

Large-scale LiDAR point cloud semantic segmentation is an important foundation for urban 3D scene modeling, road object extraction, and autonomous driving environment perception. Existing Mamba-based point cloud methods can model long-range dependencies with linear complexity. However, point cloud serialization may weaken the original neighborhood relationships, resulting in insufficient representation of slender structures such as road markings, fences, poles, and object boundaries. To address this problem, this paper proposes GradMamba-Net, a Mamba-based point cloud semantic segmentation network with local geometric enhancement. The network is collaboratively designed from three stages: before serialization, sequence encoding, and decoding recovery. First, a local gradient embedding module is used to enhance the differential geometric representation between center points and neighboring points. Second, a multi-scale spatial aggregation module is introduced during Mamba encoding to supplement the local spatial structures that are easily weakened in one-dimensional sequence modeling. Finally, an inter-channel shuffle reconstruction decoder is designed to strengthen the interaction between deep semantic features and shallow geometric details. Experiments are conducted on the Toronto-3D and WHU-Urban3D datasets. GradMamba-Net achieves an mIoU of 82.3% on the Toronto-3D dataset, outperforming 3D-UMamba by 4.1 percentage points. On the MLS-S and ALS-S scenarios of WHU-Urban3D, it achieves mIoUs of 61.2% and 64.5%, respectively. The results show that the proposed method improves the segmentation performance of slender objects and boundary regions while maintaining good inference efficiency.

Key words: LiDAR point cloud, semantic segmentation, Mamba, long-sequence modeling, local geometric enhancement, feature reconstruction

中图分类号: