顾及局部-全局特征多尺度卷积注意力网络的点云地物分类方法  

A Multi-scale Convolutional Attention Network Method for Point Cloud Object Classification Considering Local-Global Features

作　　者：隋心[1] 郝玉婷 陈志键 王长强[1] 史政旭 徐爱功[1] SUI Xin;HAO Yuting;CHEN Zhijian;WANG Changqiang;SHI Zhengxu;XU Aigong(School of Geomatics,Liaoning Technical University,Fuxin 123000,China;Department of Military Oceanography and Hydrography and Cartography,Dalian Naval Academy,Dalian 116018,China)

机构地区：[1]辽宁工程技术大学测绘与地理科学学院,阜新123000 [2]大连舰艇学院军事海洋与测绘系,大连116018

出　　处：《地球信息科学学报》2025年第2期397-410,共14页Journal of Geo-information Science

基　　金：国家自然科学基金项目(42074012);辽宁省重点研发计划项目(2020JH2/10100044);辽宁省“兴辽英才计划”项目(XLYC2002101、XLYC2008034)。

摘　　要：【目的】由于点云的非结构化和无序性,现有的深度学习点云分类网络存在局部特征和全局特征挖掘不充分并且缺乏有效的上下文特征融合的问题,难以实现地物精细分类。因此,本文提出了一种顾及局部-全局特征多尺度卷积注意力网络的点云地物分类方法。【方法】首先,针对点云的非结构性,构建局部加权图学习中心点和邻域点的位置关系,动态调整核权重,以获得更具代表性的局部特征。同时提出全局图注意力模块,考虑各点之间的全局空间分布,应对点云无序性的同时,可以有效捕获全局上下文特征,从而有效整合不同尺度信息。此外,设计自适应加权池化模块进一步实现局部和全局特征的自适应融合,最大程度提高网络的分类性能。【结果】应用开源Toronto-3D点云数据集和实测校园点云数据集验证本文方法有效性,实验结果表明,在Toronto-3D数据集本文方法的OA和MIoU分别为97.21%和85.46%,相较于Pointnet++、DGCNN、RandLA-Net、BAAF-Net和BAF-LAC等网络模型,OA提升了1.99%~8.21%,MIoU提升了3.23%~35.86%,在校园数据集本文方法的OA和MIoU分别为97.38%和85.70%,OA提升了0.58%~10.53%,MIoU提升了2.01%~32.01%。【结论】本文方法实现了复杂场景下高精度、高效率的自动化地物精细分类。[Objectives]Scene understanding based on 3D laser point clouds plays a core role in many applications such as object detection,3D reconstruction,cultural relic protection,and autonomous driving.The semantic classification of 3D point clouds is an important task in scene understanding,but due to the large amount of data,diverse targets,and large-scale differences,as well as the occlusion of buildings and trees,this task still poses challenges.The existing deep learning models for point cloud classification face several challenges due to the unstructured and disordered nature of point clouds.These challenges include inadequate extraction of local and global features and the absence of an efficient mechanism for context feature integration,making it challenging to achieve fine-grained classification of ground objects.Therefore,this study introduces a novel point cloud feature classification approach that incorporates a multi-scale convolutional attention network for both local and global features.[Methods]To address the lack of structure in point clouds,we construct a local weighted graph to model the positional relationships between central points and their neighboring points.This graph facilitates dynamic adjustments of kernel weights,enabling the extraction of more representative local features.Simultaneously,we introduce a global graph attention module to account for the overall spatial distribution of points,address the disorder of point clouds,and effectively capture global contextual features,thereby integrating information at different scales.Furthermore,we design an adaptive weighted pooling module to facilitate the seamless fusion of local and global features,thus maximizing the network's classification performance.[Results]The proposed method is evaluated using the publicly available Toronto-3D point cloud dataset and a campus point cloud dataset obtained from real measurements.We compare its performance against various network models,including Pointnet++,DGCNN,RandLA-Net,BAAF-Net,and BAF-LAC,The experime

关 键 词：深度学习 点云 地物分类 特征增强 多尺度融合 动态图卷积 注意力机制 自适应池化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]