基于高程感知多尺度图卷积网络的地物分类  

Ground object classification based on height-aware multi-scale graph convolution network

作　　者：文沛 程英蕾[1] 王鹏[1] 赵明钧 张碧秀 WEN Pei;CHENG Yinglei;WANG Peng;ZHAO Mingjun;ZHANG Bixiu(Information and Navigation College,Air Force Engineering University,Xi’an 710077,China;PLA 93575,Chengde 067000,China;PLA 93897,Xi’an 710077,China)

机构地区：[1]空军工程大学信息与导航学院,西安710077 [2]中国人民解放军93575部队,承德067000 [3]中国人民解放军93897部队,西安710077

出　　处：《北京航空航天大学学报》2023年第6期1471-1478,共8页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家自然科学基金(61773396)。

摘　　要：机载激光雷达获取的点云具有类别分布不均匀、样本高程差异大的复杂特点,现有方法难以充分识别其细粒度的局部结构。针对该问题,堆叠使用多层边缘卷积算子同时提取局部信息和全局信息,并引入高程注意力权重作为特征提取的补充,设计了一种用于机载激光雷达点云地物分类的端到端网络,提出基于高程感知多尺度图卷积网络的地物分类方法。对原始点云划分子块并采样到固定点数;采用多尺度边缘卷积算子提取多尺度局部-全局特征并进行融合,同时采用高程感知模块生成注意力权重并应用于特征提取网络;利用改进的焦点损失函数进一步解决类别分布不均问题,完成分类。采用国际摄影测量与遥感学会(ISPRS)提供的标准测试数据集对所提方法进行验证,所提方法的总体分类精度达到0.859,单类别分类精度特别是对建筑物的提取精度较ISPRS竞赛中公开的最好结果提高了4.6%。研究结果对实际应用和网络设计优化具有借鉴意义。The point cloud acquired by airborne LiDAR has the complex characteristics of uneven distribution of categories and large differences in sample elevation.Existing methods are difficult to fully identify fine-grained local structures.This paper proposes an end-to-end network for airborne LiDAR point cloud classification after employing stacked multi-layer edge convolution operators to simultaneously extract local and global information.It also introduces elevation attention weights as a supplement to feature extraction.First,the original point cloud is divided into sub-blocks and sampled to a fixed number of points.Then the multi-scale edge convolution operator is used to extract multi-scale local-global features which are merged thereafter,at the same time,the height-aware module is used to generate attention weights and applied to the feature extraction network.Finally,the improved focus loss function is used to further solve the problem of uneven distribution of categories and complete the classification.The standard test data set provided by the International Society for Photogrammetry and Remote Sensing(ISPRS)was used to verify the proposed method.Overall,85.9% of the classifications were accurate.The single-category classification accuracy,especially the roof,was increased by 4.6%than the best result published in the ISPRS competition.The research results have reference significance for practical applications and network design optimization.

关 键 词：图像处理 图卷积 深度学习 高程感知 点云分类 

分 类 号：V557.3[航空宇航科学与技术—人机与环境工程] TP391[自动化与计算机技术—计算机应用技术]