作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (7): 254-263. doi: 10.19678/j.issn.1000-3428.0062245

• 图形图像处理 • 上一篇    下一篇

点线特征融合的激光雷达单目惯导SLAM系统

崔云轩, 刘桂华, 余东应, 郭中远, 张文凯   

  1. 西南科技大学 信息工程学院, 四川 绵阳 621010
  • 收稿日期:2021-08-02 修回日期:2021-09-06 出版日期:2022-07-15 发布日期:2021-09-17
  • 作者简介:崔云轩(1998—),男,硕士研究生,主研方向为视觉SLAM、激光SLAM、三维重建;刘桂华,教授、博士;余东应、郭中远、张文凯,硕士研究生。
  • 基金资助:
    国防科工局核能开发科研项目 ([2016]1295);四川省科技厅重点研发项目(2021YFG0380)。

Lidar-Mono-Inertial SLAM System with Fusion of Point-Line Features

CUI Yunxuan, LIU Guihua, YU Dongying, GUO Zhongyuan, ZHANG Wenkai   

  1. College of Information Engineering, Southwest University of Science and Technology, Mianyang, Sichuan 621010, China
  • Received:2021-08-02 Revised:2021-09-06 Online:2022-07-15 Published:2021-09-17

摘要: 多传感器融合的SLAM系统定位精度相比单一传感器的SLAM系统更高,但在低纹理场景或退化场景下的定位精度有待提高。提出一种点线特征融合的激光雷达视觉单目惯导紧耦合SLAM系统(PL2VI-SLAM),其由点线特征融合的视觉惯导系统(PLVIS)和激光雷达惯导系统(LIS)两个子系统组成。通过PLVIS系统实现点线特征的提取与匹配,使用滑动窗口选择性地引入关键帧,并将惯性导航器件与相机紧耦合以解算位姿。LIS系统将多个约束集成到因子图中进行联合优化,其初始化状态可以作为PLVIS的初始猜测,通过扫描匹配实现激光雷达里程计,并将点云深度分别与PLVIS系统的特征点以及特征线进行关联,为视觉特征提供精确的深度值,提升定位精度。此外,两个子系统将联合进行回环检测,并对位姿进行矫正。在jackal、handled以及自制的长走廊数据集上的实验结果表明,与LVI-SAM、VINS-MDNO及LIO-SAM系统相比,该系统的定位精度更高,适用于低纹理场景及退化场景,并能满足实时性要求。

关键词: 低纹理场景, 退化场景, 点线特征提取, 线特征匹配, 特征深度关联

Abstract: The multisensor fusion Simultaneous Localization and Mapping(SLAM) system has higher localization accuracy than the single-sensor SLAM system.However, its localization accuracy in low-texture or degraded scenes needs improvement.The Point-Line with lidar-Visual-mono-Inertial tightly coupling SLAM system(PL2VI-SLAM) is proposed.This system comprises two subsystems:the Point-Line with Visual-Inertial System(PLVIS) and the Lidar Inertial System(LIS).The PLVIS subsystem first extracts and matches the point-line features.Next, this subsystem closely couples the inertial measurement unit with a camera to enhance the position by selectively introducing keyframes through sliding windows.LIS integrates multiple constraints into the factor graph joint optimization, and its initial state can be used as the initial guess of PLVIS.The lida rodometry is achieved by scan-matching, and its point cloud depth is associated with the points and feature lines of PLVIS to provide a precise depth value for visual features.These procedures further improve positioning accuracy.Finally, the two subsystems jointly conduct loop-closure to correct the position.The experimental results forthe jackal, handled, and self-made long-corridor data sets show that compared with VINS-MONO, LIO-SAM, and LVI-SAM systems, this system exhibits improved positioning accuracy, can satisfactorily handle low-texture and degraded scenes, and can meet the real-time requirements.

Key words: low texture scene, degradation scene, point line feature extraction, line feature matching, feature depth association

中图分类号: