【作者】 杨锐；

【导师】 王庆；

【作者基本信息】 东南大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 随着自主导航技术的飞速发展,激光雷达(LiDAR)和惯性测量单元(Inertial Measurement Unit,IMU)融合的定位算法已成为无人系统、机器人和自动驾驶汽车等领域的关键技术之一。但在使用激光雷达进行定位时,由于单帧激光雷达点云数据量大,通常需要进行特征筛选。现有的特征筛选方法在非结构化环境中表现不佳,可能选取到过多不稳定的特征点,造成定位精度的下降。此外,现有的定位方法大多建立离散的轨迹,在激光雷达定位过程中,只保留关键帧的位姿,这种做法丢失了轨迹的细节,导致只能使用近似的方法来估计非关键帧时刻的位姿。本文聚焦于以激光雷达为主的定位方案,深入研究了LiDAR/IMU融合定位算法,主要研究工作如下:(1)针对现有的特征筛选方法鲁棒性差的问题,提出了一种基于贪心算法的前端特征筛选方法。在前端配准的过程中,先通过自适应体素降采样对点云进行预处理,再考虑非线性优化的迭代过程,将特征选取建模成一个最大化矩阵谱特性的问题,利用贪心算法减小状态不确定度,从而筛选可靠的特征点。实验验证表明,该方法在提高定位精度方面取得了良好效果,将其集成于开源算法CT-LIO上进行测试,进行特征筛选后,在所测数据集上绝对位姿误差均方根平均降低了28%。(2)针对现有的定位算法容易受到离散化引起的误差的问题,提出一种基于哈希地图的多尺度连续时间状态估计方法。在优化阶段,将关键帧的姿态和位置作为控制点,采用B样条函数构建载体的连续时间轨迹,并通过优化控制点调整局部轨迹,从而实现载体在任意时刻的位姿查询。在地图管理阶段,在构建点云地图的过程中引入多层级哈希地图进行后端数据管理,以减少点云搜索时间,提高状态估计的精度和效率。在匹配阶段,采用多级匹配与约束构建方案,充分利用哈希地图对环境空间进行尺度划分,同时在多个尺度上构建约束关系。通过多个开源数据集上的测试表明,相比于开源算法DLIO,在所测数据集上绝对位姿误差均方根平均降低了11%,相比于开源算法FAST-LIO,在所测数据集上绝对位姿误差均方根平均降低了27%,相比于开源算法LIO-SAM,在所测数据集上绝对位姿误差均方根平均降低了34%。(3)为了验证所提出的定位方法在实际场景中的可靠性,自研了一套手持/背包式多源数据采集设备并在村镇环境下进行了实地实验,验证了所提出的连续时间状态估计方法有效地减少了定位的误差。使用自研的设备在村镇的复杂环境中的实验表明,在村镇实际场景中,相比于开源算法DLIO和FAST-LIO,本文提出算法的绝对位姿误差均方根平均降低了62%。更多还原

【Abstract】 With the rapid advancement of autonomous navigation technology,the fusion of Light Detection and Ranging(LiDAR)and Inertial Measurement Unit(IMU)for localization algorithms has become a key technology in fields such as unmanned systems,robotics,and autonomous vehicles.However,when using LiDAR for localization,feature selection is typically necessary due to the large volume of data in a single LiDAR frame.Existing feature selection methods perform poorly in unstructured environments,potentially selecting too many unstable feature points(e.g.,points in grassy areas),leading to decreased localization accuracy.Furthermore,existing localization methods mostly establish discrete trajectories,retaining only the poses of keyframes during the LiDAR localization process,thereby losing trajectory details and resorting to approximate methods to estimate poses at non-keyframe instances.The paper focuses on LiDAR-centric localization solutions and thoroughly investigates LiDAR/IMU fusion localization algorithms.The primary research contributions are as follows:(1)Addressing the problem of poor robustness in existing feature selection methods,a front-end feature selection method based on a greedy algorithm is proposed.In the front-end registration process,point cloud preprocessing is first performed using adaptive voxel downsampling,followed by consideration of a nonlinear optimization iterative process.Feature selection is modeled as a problem of maximizing matrix spectral characteristics,utilizing greedy algorithms to reduce state uncertainty and thus select reliable feature points.Experimental validation demonstrates that this method achieves good results in improving localization accuracy.Integrated into the open-source algorithm CT-LIO and tested after feature selection,the root mean square of absolute pose errors decreased by an average of 28%on the measured dataset.(2)Addressing the issue of errors caused by discretization in existing localization algorithms,a multi-scale continuous-time state estimation method based on hash maps is proposed.In the optimization phase,the poses and positions of keyframes are treated as control points,and a B-spline function is used to construct the continuous-time trajectory of the carrier.Local trajectories are adjusted through optimization of control points to achieve pose queries at arbitrary instants.In the map management phase,a multi-level hash map is introduced in the process of building point cloud maps for backend data management to reduce point cloud search time,thereby improving the accuracy and efficiency of state estimation.In the matching phase,a multi-level matching and constraint construction scheme is employed,making full use of hash maps to partition the environmental space into scales and construct constraint relationships at multiple scales.Testing on multiple open-source datasets shows that compared to the open-source algorithm DLIO,the root mean square of absolute pose errors decreased by an average of 11%on the measured dataset,compared to the open-source algorithm FAST-LIO,the root mean square of absolute pose errors decreased by an average of 27%on the measured dataset,compared to the open-source algorithm LIO-SAM,the root mean square of absolute pose errors decreased by an average of 34%on the measured dataset.(3)To verify the reliability of the proposed localization method in real-world scenarios,a self-developed handheld/backpack multi-source data collection device was built and field experiments were conducted in village environments.The effectiveness of the proposed continuous-time state estimation method in reducing localization errors was validated.Experiments in complex village environments using the self-developed device demonstrated that compared to the open-source algorithm DLIO and FAST-LIO,the root mean square of absolute pose errors decreased by 62%in actual village scenarios.更多还原