测绘学报 ›› 2021, Vol. 50 ›› Issue (11): 1594-1604.doi: 10.11947/j.AGCS.2021.20210268

• 智能驾驶环境感知 • 上一篇    下一篇

顾及室内场景特征的多线激光雷达初始定位

史鹏程1,2,3, 叶勤2, 张绍明2, 邓海峰4   

  1. 1. 自然资源部城市国土资源监测与仿真重点实验室, 广东 深圳 518034;
    2. 同济大学测绘与地理信息学院, 上海 200092;
    3. 武汉大学计算机学院, 湖北 武汉 430072;
    4. 上海华测导航技术股份有限公司, 上海 201702
  • 收稿日期:2021-05-13 修回日期:2021-10-29 发布日期:2021-12-07
  • 通讯作者: 叶勤 E-mail:yeqin@tongji.edu.cn
  • 作者简介:史鹏程(1996—),男,博士生,研究方向为点云处理,自动驾驶, 同时定位与建图。
  • 基金资助:
    国家自然科学基金(41771480);自然资源部城市国土资源监测与仿真重点实验室开放基金(KF-2019-04-015)

Localization initialization for multi-beam LiDAR considering indoor scene feature

SHI Pengcheng1,2,3, YE Qin2, ZHANG Shaoming2, DENG Haifeng4   

  1. 1. Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources, Shenzhen 518034, China;
    2. College of Surveying and Geo-Informatics, Tongji University, Shanghai 200092, China;
    3. School of Computer Science, Wuhan University, Wuhan 430072, China;
    4. Shanghai Huace Navigation Technology Co., Ltd., Shanghai 201702, China
  • Received:2021-05-13 Revised:2021-10-29 Published:2021-12-07
  • Supported by:
    The National Natural Science Foundation of China (No. 41771480);The Open Fund of Key Laboratory of Urban Land Resources Monitoring and Simulation, MNR (No. KF-2019-04-015)

摘要: 针对机器人在室内大范围场景中定位初始化技术难题,提出一种基于特征模式的定位初始化方法。首先,结合室内场景结构特征分析,探索场景中具有空间位置标示功能的稳定人工构筑物如墙壁、柱体等结构及结构组合,将其定义为特征模式,以提高场景特征表达稳健性。其次,结合多线激光雷达数据特点,提出实时数据的特征模式提取方法,对其分级管理,提高了场景特征表达效率。然后,提出一种半自动化处理方法实现点云地图特征模式提取,并采用一种高效的数据管理方案,避免在多次初始化时对地图数据重复冗余操作,提高定位效率。最后,针对各类特征模式,构建两种误差方程,结合L-M梯度下降求解方法,以地图格网击中比率作为初始化评价指标,制定自适应的匹配与配准策略,实现机器人在大尺度室内场景中的定位初始化。为了验证本文方法的可行性,使用低成本的16线激光雷达,并选取走廊、大厅、地下停车场3种典型室内场景进行试验。试验结果表明,本文方法可快速实现大尺度室内场景的定位初始化,其性能基本满足实际应用中室内机器人的定位精度与效率要求。

关键词: 自动驾驶, 室内机器人, 定位初始化, 特征模式, 激光雷达, 点云地图

Abstract: For the problem of localization initialization (LI) of robot in indoor large-scale scene, a localization initialization method based on feature pattern is proposed. Firstly, with feature analysis of indoor scene structure, the proposed method explores robust man-made structures (e.g., walls, columns and some other structures with spatial location indication function), which are defined as feature patterns to improve robustness of scene feature expression. Then, with characteristics of multi-beam light detection and ranging (LiDAR) point cloud, a feature pattern extraction method in real-time data is proposed to improve efficiency of feature expression with a hierarchical management. Next, a semi-automatic processing method is proposed to extract feature patterns from point cloud map, and an efficient data management pipeline is designed to avoid repeatedly redundant operations on map data during multiple times initialization to improve efficiency of LI. Finally, two kinds of error equations are constructed for different feature patterns. With L-M gradient descent solution and hit ratio of map grid as metric, an adaptive matching and registration strategy is proposed to accomplish LI of robot in large-scale indoor scene. In order to verify feasibility of this method, a low-cost 16-line LiDAR was used in the experiment in three typical indoor scenes i.e., corridor, hall and underground parking lot. The experimental results show that LI is accomplished quickly and accurately with proposed method in large-scale indoor scene, and it basically meets the localization accuracy and efficiency requirements of indoor robot in practical application.

Key words: autonomous driving, indoor robot, localization initialization, feature pattern, LiDAR, point cloud map

中图分类号: