基于深度学习的杆状地物分类研究  

Classification of Rod-Shaped Features Based on Deep Learning

作　　者：李佳佳 李永强[1] 杨亚伦 LI J iajia;LI Yongqiang;YANG Yalun(School of Surveying and Land Information Engineering,Henan Polytechnic University,Jiaozuo 454003,China)

机构地区：[1]河南理工大学测绘与国土信息工程学院,河南焦作454003

出　　处：《测绘地理信息》2023年第5期80-84,共5页Journal of Geomatics

基　　金：国家自然科学基金(41771491)。

摘　　要：针对城市道路场景中杆状地物分类自动化与智能化程度低的问题,提出一种基于深度学习的方法实现杆状地物的有效分类。利用改进的随机抽样一致(random sample consensus,RANSAC)算法滤除车载激光雷达(light detection and ranging,LiDAR)获取的地面点云,基于二值图像法实现点云数据分割并提取杆状地物样本,对点云数据进行归一化、下采样等处理,并将杆状地物样本按7∶3的比例制作训练集与测试集。通过改进的Point Net增强点云局部特征提取能力,将数据集送入改进的Point Net进行分类。结果表明,改进的Point Net的杆状地物分类精度为98.4%,比Point Net提高了1.8%。之后,将样本点云数量分别采样为1024、512、256、128,并制作成数据集送入网络进行训练,结果验证了该网络对杆状地物分类的鲁棒性。Aiming at the problem of low degree of automation and intelligence of rod-shaped features in the urban road scene,we propose a method to classify the rod-shaped features based on deep learning.The improved random sample consensus(RANSAC)algorithm is used to filter the ground point cloud obtained by vehicle-borne light detection and ranging(LiDAR).Based on the binary image method,we segment the point cloud data and extract the rod-shaped feature samples.The point cloud data are normalized and downsampled,and the training set and test set of the rod-shaped feature samples are made according to the ratio of 7∶3.The extraction ability of local features of point cloud is enhanced by the improved Point Net,and then,the data set is sent to the improved Point Net for classification.The results show that the accuracy of the improved Point Net for the classification of rodshaped features is 98.4%,which is 1.8%higher than that of Point Net.Then,the numbers of sample point clouds are sampled as 1024,512,256 and 128,respectively,and the datasets are made and sent to the network for training.The results verify the robustness of the network for the classification of rod-shaped features.

关 键 词：杆状地物 车载LiDAR 二值图像 归一化 下采样 Point Net 

分 类 号：P237[天文地球—摄影测量与遥感]