基于改进超点图的移动激光点云数据自动分类  

Automatic Classification of Mobile Laser Point Cloud Data Based on Improved Superpoint Graphs

作　　者：黄刚 侯燕 杨蒙蒙 HUANG Gang;HOU Yan;YANG Mengmeng(Mingli Surveying and Mapping Technology Co.,Ltd.,Beijing 100026,China;NavInfo Co.,Ltd.,Beijing 100094,China;School of Vehicle and Mobility,Tsinghua University,Beijing 100084,China)

机构地区：[1]北京明立测绘科技有限公司,北京100026 [2]北京四维图新科技股份有限公司,北京100094 [3]清华大学车辆与运载学院,北京100084

出　　处：《测绘与空间地理信息》2024年第2期9-13,共5页Geomatics & Spatial Information Technology

基　　金：国家自然科学基金(U1864203)资助。

摘　　要：针对点云数据无序、无结构、数据量大、点云密度不均匀、数据处理难度较大等问题,将超点图三维语义分割网络模型应用于移动激光点云数据自动分类领域,并提出两点优化方法:1)在PointNet网络中引入多尺度网络结构,同时获取点云的局部特征和全局特征,实现多尺度特征重用;2)使用Adam优化算法代替原有的梯度下降算法,提升深度学习性能。实验使用真实道路数据进行训练与验证,结果表明,相较于PointNet、PointNet++、PointSIFT及SPG等方法,在复杂道路情况下,精确性和均交并比有一定提高,且具有很好的鲁棒性。Aiming at the problems such as disordered,unstructured,large amount of data,uneven density of point cloud and difficulty of data processing,the 3D semantic segmentation network model of the superpoint graphs model was applied to the field of automatic classification of mobile laser point cloud data.Two optimization methods were proposed for the above problems:①A multi-scale net-work structure was introduced into PointNet network,and multi-scale feature reuse can be realized by obtained local and global fea-tures of point cloud at the same time.②The deep learning performance were improved by using Adam optimization algorithm to replace the original gradient descent algorithm.The experiment used real road data for training and verification.The experimental results show that compared with PointNet,PointNet++,PointSIFT and SPG methods,the accuracy and MIOU are improved in the case of complex roads,and the improved superpoint graphs method has good robustness.

关 键 词：图像处理 深度学习 激光点云 自动分类 语义分割 移动测量系统 

分 类 号：P231[天文地球—摄影测量与遥感]