电子测量与仪器学报

吴永豪,李胜,邹文成.动态场景中的多传感器融合SLAM算法研究[J].电子测量与仪器学报,2025,39(11):1-10

动态场景中的多传感器融合SLAM算法研究

Research on multi-sensor fusion SLAM algorithm in dynamic scenesWu YonghaoLi ShengZou Wencheng

DOI：

英文关键词:multi-sensor fusion dynamic scenes ground segmentation lidar SLAM factor graph

基金项目:江苏省科技重大专项（BG2024041）、国家自然科学基金（62373191）项目资助

作者	单位
吴永豪	南京理工大学自动化学院南京210094
李胜	南京理工大学自动化学院南京210094
邹文成	南京理工大学自动化学院南京210094

Author	Institution
Wu Yonghao	School of Automation, Nanjing University of Science and Technology, Nanjing 210094, China
Li Sheng	School of Automation, Nanjing University of Science and Technology, Nanjing 210094, China
Zou Wencheng	School of Automation, Nanjing University of Science and Technology, Nanjing 210094, China

摘要点击次数: 666

全文下载次数: 6215

中文摘要:

针对机器人在动态复杂环境下轨迹漂移以及无法建立静态地图问题，设计了一种动态点云去除的多传感器融合同步定位与建图（SLAM）算法。算法前端利用惯性测量单元（IMU）预积分实现点云畸变补偿，并采用迭代误差状态卡尔曼滤波（IESKF）算法在前端获得初始位姿估计。针对动态物体干扰，提出了一种结合地面分割和时空法向量分析的动态点云去除策略，有效剔除了动态目标的影响，保证了静态地图的全局一致性。后端基于因子图优化，融合激光惯导里程计、IMU与编码器预积分，并引入地平面因子，通过多重约束有效抑制了累积误差和Z轴漂移问题。在校园实测的复杂动态环境中，该算法相较于LeGO-LOAM、FAST-LIO和LIO-SAM主流SLAM方案，定位均方根误差(RMSE)分别降低了46.2%、49.4%和35.9%，同时有效地去除了地图中的动态点云，验证了算法的优越性，为复杂动态环境下机器人的自主导航与精确建图提供了可靠的技术支撑。

英文摘要:

To address the challenges of robot trajectory drift in dynamic and complex environments, and overcome the limitations of conventional static map construction, we propose a robust multi-sensor fusion SLAM algorithm integrated with dynamic point cloud removal. Our front-end processing employs IMU pre-integration to compensate for point cloud distortion and utilizes an iterative error state Kalman filter (IESKF) for refined initial pose estimation. Furthermore, we introduce a novel dynamic point cloud removal strategy that combines ground segmentation with spatio-temporal normal vector analysis. It effectively eliminates moving objects and preserves static structures to ensure global map consistency. On the back end, our method leverages factor graph optimization, fusing laser-inertial odometry, IMU pre-integration, and wheel encoder data to enhance trajectory estimation. In addition, we incorporate ground plane constraints to suppress cumulative errors and mitigate z-axis drift. Experimental validation in a complex campus environment demonstrates that our method significantly reduces positioning root mean square error (RMSE) by 46.2%, 49.4%, and 35.9% compared to LeGO-LOAM, FAST-LIO, and LIO-SAM, respectively. Moreover, our method successfully removes dynamic point clouds from the constructed maps, showcasing superior robustness in dynamic scenarios. These advancements provide reliable support for autonomous robot navigation and high-precision mapping in complex dynamic environments.

查看全文查看/发表评论下载PDF阅读器