杨艺, 薛晓杰, 王田,等.基于双边特征差异聚合与扩张残差的三维点云语义分割[J].河南理工大学学报（自然科学版）,2026,45(3):30-40.

YANG Y, XUE X J, WANG T,et al.Semantic segmentation of three-dimensional point cloud based on bilateral feature difference aggregation and dilated residuals[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(3):30-40.

基于双边特征差异聚合与扩张残差的三维点云语义分割

杨艺1,2, 薛晓杰1,2, 王田3, 崔科飞4, 费树岷5

1.河南理工大学 电气工程与自动化学院，河南 焦作  454003；2.河南理工大学 河南省煤矿装备智能检测与控制重点实验室，河南 焦作  454003；3.北京航空航天大学 人工智能研究院，北京  100191；4.郑州恒达智控科技股份有限公司，河南 郑州  450016；5.东南大学 自动化学院，江苏 南京 210096

摘要: 目的 点云语义分割是点云场景理解与分析的基础。为解决现有点云语义分割网络提取特征不充分，导致无法有效获取局部上下文信息且在分割复杂形状点云时表现不佳的问题，提出基于双边特征差异聚合与扩张残差的三维点云语义分割网络。  方法 首先，对局部点云的几何坐标和语义特征进行差异化编码和调整，从而使网络全面学习点与点的空间位置关系，并充分挖掘不同类型特征间的联系，增强局部上下文信息；其次，利用差异混合池化聚合突出特征和重要特征，减少信息损失，进而得到准确的单点特征表示；最后，通过扩张残差模块加深网络深度，扩大局部感受野，获得更大范围的上下文信息。  结果 在大尺度室内场景点云数据集S3DIS和室外复杂道路场景点云数据集Toronto3D上进行实验验证，结果表明，所提方法在S3DIS上的平均交并比为71.6%，平均准确率为83.2%，相较于RandLA-Net分别提升了1.6%和1.2%，在Toronto3D上的平均交并比和总体准确率分别达到了81.0%和96.8%，相比于RandLA-Net分别提升了4.7%和1.3%，整体优于所对比的点云语义分割方法，并在2个数据集的多个类别上取得最优结果。  结论 所提方法在分割复杂形状点云时，能够充分提取差异化特征信息，对点云边界做出准确划分，实现不同场景下的点云语义分割。

关键词:三维点云;深度学习;语义分割;差异聚合;残差网络

doi:10.16186/j.cnki.1673-9787.2023090059

基金项目:国家自然科学基金资助项目（61972016）

收稿日期:2025/06/27

修回日期:2025/10/13

出版日期:2026/04/28

Semantic segmentation of three-dimensional point cloud based on bilateral feature difference aggregation and dilated residuals

Yang Yi1,2, Xue Xiaojie1,2, Wang Tian3, Cui Kefei4, Fei Shumin5

1.School of Electrical Engineering and Automation， Henan Polytechnic University， Jiaozuo  454003， Henan， China；2.Henan Key Laboratory of Intelligent Detection and Control of Coal Mine Equipment， Henan Polytechnic University， Jiaozuo  454003， Henan， China；3.Institute of Artificial Intelligence， Beihang University， Beijing  100191， China；4.Zhengzhou Hengda Intelligent Control Technology Company Limited， Zhengzhou  450016， Henan， China；5.School of Automation， Southeast University， Nanjing  210096， Jiangsu， China

Abstract: Objectives Point cloud semantic segmentation was the foundation of understanding and analyzing point cloud scenes. In order to solve the problem that the existing point cloud semantic segmentation network did not extract features sufficiently， which leads to the inability to effectively obtain local context information and poor performance in segmenting complex shaped point clouds. Thus， a point cloud semantic segmentation network based on bilateral feature difference aggregation and dilated residual was proposed.  Methods Firstly， the geometric coordinates and semantic features of the local point cloud were differentially encoded and adjusted. This allowed the network to comprehensively learn the spatial relationships between points and fully explore the connections between different types of features， thereby enhancing local contextual information. Then， the salient and significant features were aggregated by using difference mixed pooling to reduce information loss and thus obtained an accurate single-point feature representation. Finally， by expanding the residual module， the network depth was deepened by the dilated residual module to expand ， the local receptive field was expanded， and obtain a larger range of contextual information was obtained.  Results Experimental validation was conducted on the large-scale indoor scene point cloud dataset S3DIS and the outdoor complex road scene point cloud dataset Toronto3D. The results indicated that the proposed algorithm had achieved the mean intersection over union of 71.6% and the mean accuracy of 83.2% on the S3DIS dataset， which was 1.6% and 1.2% improvement over RandLA-Net， respectively. On the Toronto3D dataset， the mean intersection over union and overall accuracy reached 81.0% and 96.8%， respectively， representing improvements of 4.7% and 1.3% compared to RandLA-Net. Overall， it outperformed the compared point cloud semantic segmentation methods and achieved state-of-the-artoptimal results across multiple categories in both datasets.  Conclusions In segmenting complex-shaped point clouds， our method can sufficiently extract differential feature information， accurately delineate point cloud boundaries， and achieve semantic segmentation of point clouds in different scenesscenarios.

Key words:three-dimensional point cloud;deep learning;semantic segmentation;difference aggregation;residual network