热带地理 ›› 2020, Vol. 40 ›› Issue (5): 893-902.doi: 10.13284/j.cnki.rddl.003263

• 论文 • 上一篇    下一篇

基于三维激光点云的复杂道路场景杆状交通设施语义分类

汤涌1(), 项铮2, 蒋腾平3   

  1. 1.速度时空信息科技股份有限公司,南京 210000
    2.南京国图信息产业有限公司,南京 210000
    3.江苏省地理信息资源开发与利用协同创新中心,南京 210023
  • 收稿日期:2020-03-08 修回日期:2020-04-19 出版日期:2020-09-28 发布日期:2020-10-10
  • 作者简介:汤涌(1990—),男,安徽芜湖人,工程师,主要研究方向为三维激光点云数据处理,(E-mail)1225477872@qq.com
  • 基金资助:
    江苏省研究生科研与实践创新计划项目(KYCX18_1206)

Semantic Classification of Pole-Like Traffic Facilities in Complex Road Scenes Based on LiDAR Point Cloud

Yong Tang1(), Zheng Xiang2, Tengping Jiang3   

  1. 1.Speed Space-Time Information Technology Co. , Ltd. , Nanjing 210000, China
    2.Nanjing Guotu Information Industry Co. , Ltd. , Nanjing 210000, China
    3.Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application, Nanjing 210023, China
  • Received:2020-03-08 Revised:2020-04-19 Online:2020-09-28 Published:2020-10-10

摘要:

文章提出一种完整的全自动化处理框架,基于三维激光点云数据对高速公路和城市道路场景的杆状目标进行了检测和分类,主要包括3个步骤:数据预处理、杆状目标检测和分类。其中,在数据预处理阶段,采用基于布料模拟滤波算法自动分离地面点和非地面点,然后基于欧氏距离聚类方法对非地面点进行快速聚类,以及采用迭代图割算法进一步分割目标对象;在目标检测阶段,集成先验信息、形状信息和位置导向搭建滤波器,对杆状目标进行检测;在对象分类过程中基于多属性特征,利用随机森林分类器对目标的特征进行计算和分类。并使用3个道路场景数据集进行测试,结果显示,3个数据集的整体MCC系数为95.6%,分类准确率为96.1%。这说明文章所构建方法具有较高性能。另外,该方法还可以鲁棒地检测杆状目标的重叠区域,较为适应复杂程度不同的道路场景。

关键词: 点云处理, 目标识别, 语义分类, 多层次特征, 复杂道路场景, 杆状交通设施

Abstract:

Pole-like objects are commonly occurring features on roads, and their identification in photographs is essential to the management and mapping of road information. In particular, mobile laser scanning systems comprise one of the most accurate and efficient techniques to gather road-related geospatial information. The automatic detection and classification of pole-like objects based on the point cloud data gathered by such systems enable significant reduction in costs and improvement in efficiency of mapping functions. This paper proposes a complete, fully automated processing framework that detects and classifies pole-like objects appearing in images of highways and urban streets based on the associated point cloud data. The primary workflow of the framework includes three steps: data preprocessing, detection, and classification of pole-like objects. During data preprocessing, an advanced filtration algorithm is used to automatically separate ground points from non-ground points. Subsequently, the non-ground points are clustered and overlapping clusters are further separated via a collaborative iterative graph cutting algorithm. During object detection, all available information, including shape information and position guidance, are integrated to detect pole-like objects. During object classification, the multiple aggregation levels of features and the contextual features corresponding to each object are calculated and transmitted into a random forest classifier to classify the detected pole-like structures. The proposed method was tested on three road scene datasets. The overall MCC coefficient corresponding to all three datasets was observed to be 95.6% during detection, and the overall classification accuracy corresponding to the three datasets was 96.1% during classification. Further, comparative experiments with respect to existing techniques were conducted, and the results demonstrated that the proposed method significantly improves the recognition of road facilities.

Key words: point cloud processing, object recognition, semantic segmentation, multiple aggregation level features, complex road scenes, pole-like road facilities

中图分类号: 

  • U491