Journal of Transportation Systems Engineering and Information Technology ›› 2018, Vol. 18 ›› Issue (4): 38-45.

• Intelligent Transportation System and Information Technology • Previous Articles     Next Articles

High Accuracy Vehicle Localization by Referring to Pavement Fingerprint

WANG Xiang-long, HU Zhao-zheng, LI Yi-cheng, HUANG Gang, CAI Hao   

  1. Intelligent Transport System Center, Wuhan University of Technology, Wuhan 430063, China
  • Received:2018-03-13 Revised:2018-05-06 Online:2018-08-25 Published:2018-08-27

基于路面指纹的高精度车辆定位

王相龙,胡钊政*,李祎承,黄刚,蔡浩   

  1. 武汉理工大学 智能交通系统研究中心,武汉 430063
  • 作者简介:王相龙(1988-),男,湖北武汉人,博士生.
  • 基金资助:

    国家自然科学基金/National Natural Science Foundation of China(51679181);湖北省技术创新项目重大专项/ Major Project of Technological Innovation in Hubei Province(2016AAA007);国家自然科学基金(重点)/National Natural Science Foundation of China(U1764262).

Abstract:

Visual based localization methods are widely used in vehicle localization. Since the forward and lateral views are susceptible to environmental change and traversing images of mapping data is time-consuming, this paper proposes a new down-view image-based presentation model: pavement fingerprint. The pavement fingerprint contains GPS(Global Positioning System), pavement feature and image feature points. It applies CNN (Convolutional Neural Network) and connect region to recognition the pavement feature of query image. It uses pavement feature to narrow the range of candidate map nodes which are filtered by GPS coarse localization. It improves efficiency of localization. In the experiment, the pavement fingerprint has been tested on the roads with dense pavement feature and road with sparse pavement feature, respectively. The results show that by utilizing the pavement fingerprint, the time consuming for localization reduced by 20.3% and the average localization error is 47.4 mm. This method improves the efficiency of vehicle localization and realizes vehicle localization with high precision.

Key words: intelligent transportation, pavement fingerprint, deep learning, homography, vehicle localization, localization characterization

摘要:

车辆定位广泛使用基于视觉的定位方法,针对前视图像或侧视图像易受到周围环境的影响且定位过程中需要遍历匹配地图图像导致耗时较长的问题,本文提出一种基于俯视路面图像的表征模型——路面指纹.路面指纹包含GPS(Global Positioning System),路面特征和图像特征点.该模型通过卷积神经网络(Convolutional Neural Network, CNN)结合连通区域识别待定位图像的路面特征信息,利用路面特征信息对GPS初定位筛选的地图节点进一步筛选从而提高定位效率.分别在路面特征信息密集和稀疏的路段下进行实验,实验结果表明,通过引入路面指纹使定位耗时减少20.3%,平均定位误差为47.4 mm.该方法能提高定位效率并实现高精度车辆定位.

关键词: 智能交通, 路面指纹, 深度学习, 平面单应性, 车辆定位, 定位表征

CLC Number: