引用本文: 姚二亮, 宋海涛, 赵婧, 范晓婧. 考虑点云结构和表观信息的激光雷达－惯性SLAM算法[J]. 机器人, 2024, 46(4): 436-449.
YAO Erliang, SONG Haitao, ZHAO Jing, FAN Xiaojing. LiDAR-inertial SLAM Algorithm Based on Point Cloud Structure and Appearance[J]. ROBOT, 2024, 46(4): 436-449.
 Citation: YAO Erliang, SONG Haitao, ZHAO Jing, FAN Xiaojing. LiDAR-inertial SLAM Algorithm Based on Point Cloud Structure and Appearance[J]. ROBOT, 2024, 46(4): 436-449.

## LiDAR-inertial SLAM Algorithm Based on Point Cloud Structure and Appearance

• 摘要: 为提高移动机器人在卫星拒止环境下的精确导航能力，克服大范围运动场景下的定位误差累积问题，提出一种考虑点云结构和表观信息的激光雷达－惯性SLAM（同时定位与地图创建）算法。该算法主要包含2个部分：考虑点云结构的激光雷达－惯性里程计以及基于点云表观的闭环检测与优化。激光雷达－惯性里程计线程中，通过IMU（惯性测量单元）信息预测导航状态量，构建了考虑点云结构的直接对齐观测方程，并采用迭代误差扩展卡尔曼滤波器实时地更新导航状态量。闭环检测与优化线程中，根据点云表观差异、相对运动位姿和时间约束判定关键帧，而后基于点云表观筛选得到候选闭环关键帧，并根据表观匹配距离和2维距离对候选闭环关键帧进行排序，最后确认闭环关键帧并构建位姿图，实现全局位姿优化和地图调整。在西安世博园开展的自主导航实验表明，该算法能够实时完成自主导航，并在环形运动中准确检测闭环，有效完成闭环优化，轨迹起点和终点的平均误差仅为0.07 m。此外，进一步测试了激光雷达退化环境下的多模态组合导航切换方法，该方法有效提升了自主导航的可靠性。

Abstract: To enhance precise navigation capabilities of mobile robots operating in the satellite-denied environments and to mitigate accumulated positioning error in the large-scope motion scenarios, a LiDAR (light detection and ranging) -inertial SLAM (simultaneous localization and mapping) algorithm based on point cloud structure and appearance is proposed. The proposed algorithm consists of two parts: the LiDAR-inertial odometry considering the point cloud structure, and the loop closure detection and optimization based on the point cloud appearance. In the part of LiDAR-inertial odometry, navigation states are predicted by IMU (inertial measurement unit) information, a direct alignment observation equation is constructed with the point cloud structure considered, and navigation states are updated in real-time by an iterative error extended Kalman filter. In the part of loop closure detection and optimization, keyframes are identified by evaluating discrepancies in point cloud appearance, relative motion poses and time constraints. Then, candidate loop keyframes are screened out based on point cloud appearances. The candidate loop keyframes are subsequently sorted according to the appearance matching distances and the two-dimensional distances. Finally, the loop keyframe is confirmed and the pose graph is constructed for the global pose optimization and the map adjustment. Autonomous navigation experiments in Xi'an Expo Park show that autonomous navigation is realized in real-time, the loops are accurately detected in the loop motion, and the loop optimization is effectively completed by the proposed algorithm. The average error between the starting and ending points of trajectories is only 0.07 m. Moreover, a multimodal integrated navigation switching method in the LiDAR degradation environments is further discussed, which effectively improves the reliability of autonomous navigation.

/

• 分享
• 用微信扫码二维码

分享至好友和朋友圈