引用本文:陶倩文,胡钊政,孙勋培,万金杰,陈琪莉.基于孪生网络的场景编码与智能车多尺度激光定位[J].控制理论与应用,2025,42(3):521~530.[点击复制]
TAO Qian-wen,HU Zhao-zheng,SUN Pei-xun,WAN Jin-jie,CHEN Qi-li.Scene coding for multi-scale LiDAR-based localization of intelligent vehicle with Siamese network[J].Control Theory and Technology,2025,42(3):521~530.[点击复制]
基于孪生网络的场景编码与智能车多尺度激光定位
Scene coding for multi-scale LiDAR-based localization of intelligent vehicle with Siamese network
摘要点击 25  全文点击 2  投稿时间：2023-04-01  修订日期：2025-03-03
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2024.30182
  2025,42(3):521-530
中文关键词  智能车  激光定位  地图匹配  孪生网络
英文关键词  intelligent vehicle  LiDAR localization  map matching  Siamese network
基金项目  国家重点研发计划项目(2022YFB2502904), 湖北省重点研发计划项目(2022BAA082), 武汉市人工智能创新专项项目(2022010702040064), 重庆 市科技创新重大研发项目(CSTB2022TIAD–STX0003)资助.
作者单位E-mail
陶倩文 武汉理工大学  
胡钊政* 武汉理工大学 智能交通系统研究中心；武汉理工大学重庆研究院 zzhu@whut.edu.cn 
孙勋培 武汉理工大学 智能交通系统研究中心  
万金杰 武汉理工大学 智能交通系统研究中心  
陈琪莉 武汉理工大学 智能交通系统研究中心  
中文摘要
      基于激光点云地图的智能车定位方法存在地图数据量大、匹配准确率低等问题,为此,本文提出了基于孪生网络的场景编码与智能车多尺度激光定位方法.首先,构建了基于场景编码的激光极化地图,该地图以节点形式表征,每个节点均包含点云极化图、点云场景编码以及全局位姿;然后,提出了基于激光极化地图的智能车多尺度定位方法,包括基于GPS与运动模型的粗定位、基于直方图滤波的节点级定位和基于快速的GICP算法的度量级定位;最后,使用现场采集的校园道路数据集和公共KITTI数据集对所提出的算法进行测试验证.实验结果表明,基于所构建地图的节点级定位准确率分别为99.6%和96.9%,平均定位误差分别为0.34 m和0.21 m,且对不同类型的激光雷达传感器和不同的环境具有较强的鲁棒性.
英文摘要
      Intelligent vehicle localization methods based on LiDAR point cloud maps have problems such as large amounts of map data and low matching accuracy, hence, a scene coding method and a multi-scale LiDAR-based localization method of intelligent vehicles based on Siamese network are proposed. Firstly, a polarized LiDAR map is constructed based on scene coding, which is represented with nodes, and each node contains a polarized LiDAR image, a point cloud scene coding, and a global pose. Secondly, a multi-scale localization of intelligent vehicles is readily realized based on the polarized LiDAR map, which contains coarse localization based on global positioning system (GPS) and motion model, node-level localization based on a histogram filter, and metric-level localization based on GICP algorithm. Finally, the proposed method is verified by the campus road dataset and the public KITTI dataset. The experimental results demonstrate that the accuracies of the node-level localization are 99.6% and 96.9%, the average localization errors are 0.34 m and 0.21 m, and the proposed method has strong robustness to different types of LiDAR sensors and different environments.