汽车工程 ›› 2022, Vol. 44 ›› Issue (2): 190-198.doi: 10.19562/j.chinasae.qcgc.2022.02.005

所属专题: 智能网联汽车技术专题-感知&HMI&测评2022年

• • 上一篇    下一篇

基于2阶HMM的智能车视觉地图定位方法

周哲1,胡钊政1,2(),王志强1,肖汉彪1   

  1. 1.武汉理工大学智能交通系统研究中心,武汉  430063
    2.武汉理工大学重庆研究院,重庆  401120
  • 修回日期:2021-11-04 出版日期:2022-02-25 发布日期:2022-02-24
  • 通讯作者: 胡钊政 E-mail:zzhu@whut.edu.cn
  • 基金资助:
    国家自然科学基金(U1764262);国家重点研发计划(2018YFB1600801);重庆市自然科学基金(cstc2020jcyj-msxmX0978);武汉市科技局技术创新项目(2020010601012165)

Visual Map Matching Method for Intelligent Vehicles Based on Second-order HMM

Zhe Zhou1,Zhaozheng Hu1,2(),Zhiqiang Wang1,Hanbiao Xiao1   

  1. 1.Intelligent Transport System Center,Wuhan University of Technology,Wuhan 430063
    2.Chongqing Research Institute of Wuhan University of Technology,Chongqing 401120
  • Revised:2021-11-04 Online:2022-02-25 Published:2022-02-24
  • Contact: Zhaozheng Hu E-mail:zzhu@whut.edu.cn

摘要:

本文中针对视觉地图匹配问题,将视觉地图匹配问题转化为基于图像序列的最优视觉地图节点匹配问题,并提出基于2阶隐马尔科夫模型(hidden Markov model,HMM)的视觉地图匹配方法。在该模型中,状态变量被定义为高精度视觉地图节点,查询图像被定义为观测数据。在状态转移模型中,引入2阶模型对短时间车辆运动进行匀速运动建模,与传统的1阶HMM相比,可以提高模型的适用性与准确性。提出利用全局图像特征建立查询图像与地图节点之间的匹配关系,并从匹配的汉明距离建立发射概率模型,可有效提高地图匹配的效率。最后,通过前向算法来求解最优匹配的地图节点。为了验证算法的性能,分别在封闭工业园区、开放道路和KITTI公开数据集对算法进行验证。实验结果表明:2阶HMM模型能够有效融合车辆运动信息和图像信息,提高匹配的稳定性和精确度,算法性能明显优于传统的基于单帧匹配和序列匹配算法。

关键词: 智能车, 视觉定位, 2阶隐马尔可夫模型, 视觉地图

Abstract:

In this paper, the visual map matching problem is transformed into the optimal visual map node matching problem based on image sequence and a method for visual map matching method based on second-order HMM (Hidden Markov Model) is proposed. In this model, state variables are defined as high precision visual map nodes, and query images are defined as observation variables. In the state transition model, the second-order model is introduced to model the uniform motion of vehicles within a short period of time. Compared with the traditional first-order HMM, the second-order HMM method is more applicable and precise. The paper proposes to use the global image features to establish the matching relationship between the query image and the map nodes, and establish the transmission probability model from the matching Hamming distance, which can effectively improve the efficiency of map matching. Finally, the optimal matching map nodes are obtained by forward algorithm. The performance of the algorithm is verified in closed industrial park, open roads, and the public KITTI datasets, respectively. The experimental results show that the proposed second-order HMM model can effectively integrate vehicle motion information and image information, improve the stability and accuracy of matching, and the performance of the proposed algorithm outperforms traditional single frame matching and sequence matching algorithms.

Key words: intelligent vehicle, vision-based positioning, second-order Hidden Markov Model, visual map