融合CNN和MRF的激光点云层次化语义分割方法  

A LiDAR point cloud hierarchical semantic segmentation method combining CNN and MRF

作　　者：蒋腾平 王永君[2,4,5] 张林淇 梁冲 孙剑 JIANG Tengping;WANG Yongjun;ZHANG Linqi;LIANG Chong;SUN Jian(Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources, Shenzhen 518034, China;Key Laboratory of Virtual Geographic Environment, Ministry of Education, Nanjing Normal University, Nanjing 210093, China;State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China;Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application, Nanjing 210023, China;State Key Laboratory Cultivation Base of Geographical Environment Evolution, Nanjing Normal University, Nanjing 210093, China)

机构地区：[1]自然资源部城市土地资源监测与仿真重点实验室,广东深圳518034 [2]南京师范大学虚拟地理环境教育部重点实验室,江苏南京210093 [3]武汉大学测绘遥感信息工程国家重点实验室,湖北武汉430079 [4]江苏省地理信息资源开发与利用协同创新中心,江苏南京210023 [5]地理环境演化模拟国家重点实验室建设培育点,江苏南京210093

出　　处：《测绘学报》2021年第2期215-225,共11页Acta Geodaetica et Cartographica Sinica

基　　金：自然资源部城市土地资源监测与仿真重点实验室开放基金资助课题(KF-2018-03-070);国家自然科学基金(41771439);国家重点研发计划(2016YFB0502304);江苏省研究生科研与实践创新计划项目(KYCX18_1206)。

摘　　要：三维点云语义分割的结果包含着对场景中多个目标的识别,是三维场景信息提取的重要环节,在智慧城市等多个领域扮演关键角色。由于三维激光点云数据量庞大、场景复杂性高等问题,大多数现有方法只能以相对较低的识别率提取有限类型的对象。本文提出了一种在三维激光点云场景中结合残差学习和马尔可夫随机场(MRF)优化的层次化多类型目标自动提取框架。该框架首先将点云滤波为地面点和非地面点;然后从非地面点中提取建筑物以降低场景复杂度;接着基于现有深度模型引入残差学习模块对剩下点云进行逐点分类;最后采用马尔可夫随机场进行后处理和分类结果优化,以提高激光点云语义分割的准确率。对3个室外大规模点云场景分别进行的试验结果表明,本文方法可以对多种类型的点云场景进行有效语义分割,每个数据集的3项指标(召回率、精确度和F1值)分别为(94.6%、96.8%、95.7%)、(88.5%、90.5%、89.2%)和(95.3%、95.2%、95.3%)。此外,与现有较前沿方法相比,本文方法显著提高了语义分割性能。The result of point cloud semantic segmentation includes the recognition of multiple objects in the scene,which is an important part of 3D scene information extraction.It also plays a key role in many fields such as smart cities.Since the large amount of data and high scene complexity,however,most existing methods can only extract a limited type of objects with a relatively low recognition rate from laser scanning data.This paper proposes a hierarchical multiple object automatic extraction framework combining residual learning and Markov random field(MRF)optimization in a 3D LiDAR point cloud scene.The framework first filters raw data into ground points and non-ground points;buildings are extracted from non-ground points to reduce scene complexity;then the residual learning network is adopted to pointwise classify the remaining point clouds;finally,a MRF-based optimization is used for post-processing and improving the accuracy of point cloud classification.In order to evaluate the effectiveness and robustness of the proposed method,experiments were performed on three outdoor large-scale point cloud scenarios.Experimental results show that the proposed method can perform effective semantic segmentation on various types of point cloud scenes,with(94.6%,96.8%,95.7%),(88.5%,90.5%,89.2%)and(95.3%,95.2%,95.3%),respectively.In addition,compared with existing advanced methods,it is shown that our method significantly improves the performance of semantic segmentation.

关 键 词：激光点云 语义分割 层次化提取 残差学习 马尔可夫随机场(MRF) 

分 类 号：P237[天文地球—摄影测量与遥感]