地球信息科学学报 ›› 2019, Vol. 21 ›› Issue (7): 994-1008.doi: 10.12082/dpxxkx.2019.180697

• 地球信息科学理论与方法 • 上一篇    下一篇

结合SVM与图匹配的车载激光点云道路标线识别

方莉娜1,2,3(), 黄志文1,2,3, 罗海峰1,2,3, 陈崇成1,2,3   

  1. 1. 福州大学地理空间信息技术国家地方联合工程研究中心, 福州 350002
    2. 空间数据挖掘与信息共享教育部重点实验室, 福州 350002
    3. 福建省空间信息工程研究中心,福州 350002
  • 收稿日期:2018-12-28 修回日期:2019-03-13 出版日期:2019-07-25 发布日期:2019-07-31
  • 作者简介:作者简介:方莉娜(1983-),女,广西桂林人,博士,助理研究员,主要从事激光雷达数据处理与三维重建。E-mail: <email>fangln@fzu.ded.cn</email>
  • 基金资助:
    国家自然科学基金青年基金项目(41501493);福建省自然科学基金项目(2017J01465);中国博士后科学基金项目(2017M610391);福建省教育厅中青年教师科研项目(JAT160078)

Integrating SVM and Graph Matching for Identifying Road Markings from Mobile LiDAR Point Clouds

Li'na FANG1,2,3,*(), Zhiwen HUANG1,2,3, Haifeng LUO1,2,3, Chongcheng CHEN1,2,3   

  1. 1. National Engineering Research Centre of Geospatial Information Technology, Fuzhou University, Fuzhou 350002, China
    2. Key Laboratory of Spatial Data Mining and Information Sharing of Ministry of Education, Fuzhou University, Fuzhou 350002, China
    3. Spatial Information Research Center of Fujian Province, Fuzhou University, Fuzhou 350002, China
  • Received:2018-12-28 Revised:2019-03-13 Online:2019-07-25 Published:2019-07-31
  • Contact: Li'na FANG E-mail:fangln@fzu.ded.cn
  • Supported by:
    National Natural Science Foundation of China, No.41501493;National Natural Science Foundation of Fujian Province, China, No.2017J01465;China Postdoctoral Science Foundation, No.2017M610391;Science Foundation of Education Commission of Fujian Province, China, No.JAT160078

摘要:

本文提出一种基于SVM与图匹配相结合的车载激光点云道路标线识别方法。该方法基于标线点云分割对象,利用Hu不变矩、实心形状上下文(SSC)、最小外包矩形(MBR)面积和延展度构建形状特征向量,采用SVM进行道路标线粗分类。针对粗分类结果,构建能够精确描述空间语义信息(如局部区域内标线间的排列、方向、距离)的图结构,通过图匹配方法优化粗分类结果,完成直行箭头、人行横道预告标识线、单向转向箭头、双向转向箭头、虚线型标线、斑马线共六类道路标线的精确识别。本文实验采用4份不同场景车载激光点云数据,实验结果中6类标线分类的准确率分别达100%、100%、94.12%、100%、94.94 %、99.25%,召回率分别达100%、100%、88.89%、100%、98.21%、99.00%,F1-Measure值分别达100%、100%、91.43%、100%、96.59%、99.12%。结果表明,本文方法能实现多类标线对象的精确识别,并对形状相似标线(如直行箭头、虚线型标线与斑马线)的区分具有较强稳健性。

关键词: 车载激光点云, 道路标线识别, 形状特征, SVM, 图结构, 图匹配

Abstract:

This paper presented a novel method for identifying road markings from mobile LiDAR point clouds by integrating Support Vector Machine (SVM) and graph matching. Firstly, the road surface point cloud was extracted using the scanline method, and then, was used to generate independent marking objects by a marking segmentation method combined with intensity correction. Next, hierarchical classification was conducted to identify smaller size markings as the basic objects for further processing. Considering the shape varieties of different types of marking objects, Hu invariant moments, Solid Shape Context (SSC), the area of Minimum Bounding Rectangle (MBR), and extensibility were extracted to construct the marking shape feature vector. Subsequently, with the shape feature vector of the above-mentioned samples, a training sample set was manually established to fit SVM model parameters. In the classification section, the SVM model was conducted for the preliminary classification, where there were situations including cross misclassification of markings with similar shapes and ambiguity between dotted marking and zebra crossing. Building upon the graph structure including certain types of interactive relationship (e.g., arrangement of road markings, direction relationship, and distance between markings in the local area) and the shape feature of markings, we used the inherent characteristics and spatial neighborhood information of objects to synthetically describe the geometric feature and spatial semantic information of road markings. Since the partial absence of markings' semantic structure, we developed an inexact graph matching method based on the graph structure of markings, which could optimize the preliminary classification result. The refined classification results of all six types of road markings included the straight arrow, the crosswalk warning line, the one-way steering arrow, the two-way steering arrow, the dotted marking, and the zebra crossing. To verify the validity of the proposed method, we conducted experiments using four test data sets acquired from different MLS systems. The verification results show that the six marking types have a respective precision of 100%, 100%, 94.12%, 100%, 94.94%, and 99.25%, and a respective recall rate of 100%, 100%, 88.89%, 100%, 98.21%, and 99%, and a respective F1-Measure value of 100%, 100%, 91.43%, 100%, 96.59%, and 99.12%. The experimental results demonstrate that the proposed method could accurately identify multi-class road markings. In conclusion, our algorithm is superior and robust for extracting road markings in even complicated cases where road markings may exhibit significant shape differences or high similarity (e.g., the straight arrow and linear marking, the dotted line and zebra crossing).

Key words: mobile point clouds, road marking recognition, shape features, SVM, graph structure, graph matching