基于隐式神经地图的LiDAR/IMU匹配定位算法  

LiDAR/IMU matching localization algorithm based on implicit neural map

作　　者：高旺 赵恒 刘宏 潘树国[1] 黄飞璇 Gao Wang;Zhao Heng;Liu Hong;Pan Shuguo;Huang Feixuan(School of Instrument Science and Engineering,Southeast University,Nanjing 210096,China)

机构地区：[1]东南大学仪器科学与工程学院,南京210096

出　　处：《仪器仪表学报》2024年第12期74-84,共11页Chinese Journal of Scientific Instrument

基　　金：国家重点研发计划(2021YFB3900804)项目资助。

摘　　要：针对现有匹配定位算法依赖高内存且稠密点云地图的问题,提出基于隐式神经地图的激光雷达/惯性测量单元(LiDAR/IMU)融合匹配定位算法。首先,利用浅层感知机预测有向距离场构建轻量化及高分辨率的隐式神经地图;其次,通过点到隐式神经模型配准方法,实现基于轻量化隐式神经地图的低频状态估计。同时,针对单一激光雷达隐式配准过程中难以应对激进运动的问题,引入IMU预积分方法为隐式配准提供状态估计预测值,以减少配准过程的迭代次数;最后,基于因子图融合激光里程计因子和IMU预积分因子,实现鲁棒的高频状态估计。在KITTI数据集、实测室内走廊及室外校园环境下的实验结果表明,所构建的隐式神经地图内存占用相较于传统点云地图可减少87%,实现更轻量化的地图表示;在KITTI数据集中,提出的激光雷达隐式配准算法定位精度相比传统NDT算法提高了43.4%;引入IMU后,该融合算法定位精度相比NDT-IMU算法提高了60%;在室内和室外校园实测数据中,也验证了算法在小场景下的厘米级实时定位能力。同时分析得出,IMU的引入使得隐式神经地图配准耗时大幅减少,显著提升了地图匹配定位的实时性。Aiming at the problem that existing matching localization algorithms rely on high-memory and dense point cloud map,a light detection and ranging(LiDAR)/inertial measurement unit(IMU)matching localization algorithm based on an implicit neural map is proposed.Firstly,a lightweight and high-resolution implicit neural map is constructed using a shallow perceptron to predict the signed distance field.Secondly,low-frequency state estimation based on the lightweight implicit neural map is realized through a point and implicit neural model registration method.Meanwhile,to address the challenge of handling aggressive motion during single LiDAR implicit registration,an IMU pre-integration method is introduced.This provides predictive state estimation for implicit registration,reducing the number of iterations required during the registration process.Finally,robust high-frequency state estimation is realized by fusing the LiDAR odometry factor and IMU pre-integration factor based on the factor graph.Experimental results in the KITTI dataset,as well as in real-world indoor corridor and outdoor campus environments,demonstrate the effectiveness of the proposed algorithm.The memory usage of the implicit neural map is reduced by 87%compared to traditional point cloud maps,enabling a more lightweight map representation.In the KITTI dataset,the proposed LiDAR implicit registration algorithm improves the positioning accuracy by 43.4%compared to the traditional normal distributions transform(NDT)algorithm.Furthermore,the fusion algorithm,which incorporates IMU data,achieves a 60%improvement in positioning accuracy compared to the NDT-IMU algorithm.The centimeter-level real-time positioning capability of the proposed algorithm in small scenes is also verified in both practical indoor and outdoor campus.Meanwhile,analysis reveals that the integration of IMU significantly reduces the computational time required for implicit neural map registration,effectively enhancing real-time localization performance.Aiming at the problem that exist

关 键 词：隐式神经表征 轻量化地图 先验地图 匹配定位 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置] TH76[自动化与计算机技术—控制科学与工程]