扶晓玥，李少达，付业平，等. 基于局部双注意力的三维点云语义分割算法[J].河南理工大学学报(自然科学版), doi: 10.16186/j.cnki.1673-9787.2024110038.

FU X Y，LI S D，FU Y P，et al. A 3D point cloud semantic segmentation algorithm based on local dual attention[J]. Journal of Henan Polytechnic University( Natural Science), doi: 10.16186/i.enki.1673-9787.2024110038.

基于局部双注意力的三维点云语义分割算法 (网络首发)

扶晓玥 ¹，李少达 ¹，付业平 ²，杨容浩 ¹，谭骏祥 ¹，胡古月 ¹，谭骏祥¹，兰青龙 ¹

（1.成都理工大学 地球与行星科学学院，四川 成都 610059；2. 四川省自然资源测绘地理信息有限责任公司，四川 成都 610059）

摘要: 目的 为了提升三维点云语义分割任务中对距离较近、几何结构相似的不同语义类别地物的识别准确性和分割完整性，本文对点云语义分割方法中的局部特征提取与上下文信息融合机制进行研究，方法 提出了一种基于局部双注意力的三维点云语义分割方法（Dual Local Attention-Residual Network, DLA-ResNet）。该方法通过设计基于相对距离的编码模块，建立邻域内点之间的相对位置关系；引入基于邻域增强的自注意力模块，增强邻域内不同语义类别的特征表达；改进局部注意力池化模块，自适应地聚合突出特征；并利用残差连接结构融合上下文信息，获得全局特征，从而提高网络对具有相似特征但不同类别地物的识别能力。为验证方法的有效性，本文在自建数据集CDUT-Road以及公开数据集Toronto3D和S3DIS上进行了实验。结果 实验结果表明，本文提出的DLA-ResNet方法在CDUT-Road、Toronto3D和S3DIS数据集上的平均交并比（mIoU）分别达到了67.4%、78.8%和73.7%，相较于RandLA-Net分别提高了5.6%, 1.1%和3.7%。结论 本文提出的DLA-ResNet方法通过引入局部双注意力机制和残差连接结构，提升了模型对复杂场景中点云数据的处理能力，显著提升了分割精度，在室内和室外场景的点云数据上均表现出良好的分割效果，且在不同数据集上均表现稳健。

关键词: 车载LiDAR点云；语义分割；深度学习；注意力机制；邻域增强

中图分类号：TP391.41  　

doi: 10.16186/j.cnki.1673-9787.2024110038

基金项目: 国家自然科学基金资助项目（42401425）；四川省科技计划项目（应用基础研究）（2021YJ0369）

收稿日期：2024-11-20

修回日期：2025-02-25

网络首发日期：2025-04-15

A 3D point cloud semantic segmentation algorithm based on local dual attention (Online)

FU Xiaoyue¹，LI Shaoda¹，FU Yeping²，YANG Ronghao¹，Tan Junxiang¹，Hu Guyue¹，

Lan Qinglong¹

（1. College of Earth and Planetary Sciences，Chengdu University of Technology，Chengdu， 610059，Sichuan，China；2. Sichuan Natural Resources Mapping and Geographic Information Co.， Ltd.，Chengdu 610059，Sichuan，China）

Abstract: Objectives To improve the recognition accuracy and segmentation completeness of 3D point cloud semantic segmentation for objects that are spatially close and share similar geometric structures but belong to different semantic categories, this study investigates the mechanisms of local feature extraction and contextual information fusion in point cloud semantic segmentation methods. Methods It proposes a novel 3D point cloud semantic segmentation method based on dual local attention, termed Dual Local Attention-Residual Network (DLA-ResNet). The method designs a relative distance-based encoding module to establish the relative positional relationships among points within a neighborhood, introduces a neighborhood-enhanced self-attention module to enhance the feature representation of different semantic categories within the neighborhood, improves the local attention pooling module to adaptively aggregate salient features, and utilizes a residual connection structure to fuse contextual information and obtain global features, thereby enhancing the network's ability to distinguish objects with similar features but different categories. To validate the effectiveness of the proposed method, experiments were conducted on the self-constructed CDUT-Road dataset as well as the public Toronto3D and S3DIS datasets. Results The experimental results show that the proposed DLA-ResNet method achieves mean Intersection over Union (mIoU) values of 67.4%, 78.8%, and 73.7% on the CDUT-Road, Toronto3D, and S3DIS datasets, respectively, representing improvements of 5.6%, 1.1%, and 3.7% compared to RandLA-Net. Conclusions The proposed DLA-ResNet method enhances the model's ability to process point cloud data in complex scenes by introducing dual local attention mechanisms and residual connection structures, significantly improving segmentation accuracy. The method demonstrates excellent segmentation performance on both indoor and outdoor point cloud data and exhibits robust performance across different datasets.

Key words：MLS LiDAR point cloud; semantic segmentation; deep learning; attention mechanism; neighborhood enhanced

CLC: TP391.41