【作者】 陈晓宇；

【导师】 王庆； 张鹤年；

【作者基本信息】 东南大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 为了使得四足机器人能够实现较好的定位与建图效果,针对目前四足机器人的抖动频率较高特性以及在树木较多等复杂环境下单一传感器无法完成定位与建图问题,本文提出了一种基于Li DAR/IMU/GNSS融合的四足机器人定位与建图技术,并在室内外场景下通过实验验证了所提方法的有效性。本文的主要研究内容如下:(1)针对四足机器人载体抖动频率较高而造成点云畸变以及目前算法中没有充分考虑当前帧的结构特征问题,本文提出了一种面特征点分类的方法。该方法将激光点云的所有点视为面点,并在点云阵列上提取面特征明显点,即将点云分为面特征明显和不明显点。具体步骤是,首先利用基于B样条标定方法得到激光雷达与IMU的外参标定矩阵;之后使用IMU传感器信息进行位姿推算去除激光点云畸变;建立点云阵列进行离群点剔除,同时在点云阵列上搜索每个点周围的四个点构成邻域拟合平面,当拟合的平面效果较好时邻域内点为面特征明显点;最后,将面特征分类方法应用于FAST-LIO2算法进行比较,在所测数据集上绝对位姿误差平均降低了12.20%,相对位姿误差平均降低了10.57%。(2)针对四足机器人在复杂环境下容易出现面特征点误匹配以及在局部地图中搜索最近邻点无法保证较好拟合平面问题,本文提出了一种优先级点云匹配的方法。一级优先是当前匹配点是面特征明显点且搜索到的5个点不共线;二级优先是当前匹配点为面特征不明显点且搜索到的5个点不共线;三级优先是当前匹配特征点是面特征明显点且搜索到的5个点共线;四级优先是面特征不明显点且搜索到的5个点共线。在优先级的基础上融合强度信息重新定义了权重因子,以实现多策略的自适应加权。将该方法应用于FAST-LIO2算法,并与FAST-LIO2等算法进行比较。实验表明,优先级点云匹配方法相比于FAST-LIO2在所测数据集上绝对位姿误差平均降低了26.15%,相对位姿误差平均降低了3.5%。(3)针对目前以滤波方法为后端优化的算法中常把测量噪声R设置为常数,无法实时更新测量噪声的问题,本文在以上改进的基础上提出自适应调整测量噪声的迭代误差状态卡尔曼滤波算法,通过引入点云匹配残差修正测量噪声,并在公开数据集上与FAST-LIO2、面特征点分类方法、优先级点云匹配方法等进行比较。实验表明,相比于FAST-LIO2在所测数据集上绝对位姿误差平均降低了24.23%,相对位姿误差平均降低了7.5%。(4)针对激光雷达与IMU融合的累计误差问题,本文提出了半紧耦合的方法融合GNSS信息修正累计误差,通过剔除GNSS信号质量差的信息后以角点方式加入到位姿图中修正激光与IMU融合位姿信息。并在室内外环境运行四足机器人进行测试,并利用GNSS RTK信息与融合GNSS前后的改进算法相对比,实验表明融合GNSS信息之后相比于没有融合GNSS信息的改进算法绝对位姿误差平均降低了43.9%。更多还原

【Abstract】 In order to achieve better positioning and mapping effects for quadruped robots,a quadruped robot positioning and mapping technology based on Li DAR/IMU/GNSS fusion is proposed to address the high vibration frequency of quadruped robots and the inability of a single sensor to complete positioning and mapping in complex environments such as trees.The effectiveness of the proposed method is verified through experiments in indoor and outdoor scenarios.The main research content of this article is as follows:(1)A method for surface feature point classification is proposed to address the point cloud distortion caused by the high frequency of jitter in quadruped robots and the insufficient consideration of the structural features of the current frame in current algorithms.This method treats all points of the laser point cloud as planar points,and extracts obvious point features on the point cloud array.So the point cloud is divided into obvious and inconspicuous point features.Firstly the B-spline calibration method is used to obtain the external parameter calibration matrix of Li DAR and IMU;Then,the IMU sensor information for pose estimation is used to remove laser point cloud distortion;a point cloud array for outlier removal is established and searched for four points around each point on the point cloud array to form a neighborhood fitting plane.When the fitting plane effect is good,the points in the neighborhood are obvious points with surface features;Finally,the surface feature classification method is applied to the FAST-LIO2 algorithm for comparison,and the average absolute pose error is reduced by 12.20% and the average relative pose error is reduced by 10.57%on the measured dataset.(2)A priority point cloud matching method is proposed to address the issue of quadruped robots prone to mismatched surface feature points in complex environments and the inability to ensure good plane fitting by searching for nearest neighbors in local maps.The first priority is that the current matching point is a point with obvious planar features and the 5 searched points are not collinear;The second level priority is that the current matching point is a point with indistinct planar surfaces and the 5 points searched are not collinear;The third level priority is that the current matching feature point is a prominent point in the planar feature and the 5 points are collinear;The fourth priority is to search for 5 points that are collinear and has indistinct surface features.On the basis of priority,the weight factor is redefined by fusing strength information to achieve adaptive weighting of multiple strategies.Apply this method to the FAST-LIO2 algorithm and compare it with algorithms such as FAST-LIO2.The experiment shows that the priority point cloud matching method reduces the absolute pose error by an average of 26.15% and the relative pose error by an average of 3.5% on the measured dataset compared to FAST-LIO2.(3)In response to the problem that measurement noise R is often set as a constant in algorithms optimized based on filtering methods,which cannot update measurement noise in real-time,a Kalman filtering algorithm is proposed that adaptively adjusts the iterative error state of measurement noise based on the above improvements.The measurement noise is corrected by introducing point cloud matching residuals.The method is combined with FAST-LIO2,and compared to surface feature points classification method,and priority point cloud matching methods,etc.The experiment shows that compared to FAST-LIO2,the absolute pose error on the measured dataset is reduced by an average of 24.23%,and the relative pose error is reduced by an average of 7.5%.(4)A semi-tightly coupled method for fusing GNSS information is proposed to address the cumulative error issue in the fusion of Li DAR and IMU.The method corrects the cumulative error by removing information about poor GNSS signal quality and adds corner points to the pose map to correct the fusion pose information of the Li DAR and IMU.The quadruped robot is tested in indoor and outdoor environments,and the improved algorithm before and after GNSS fusion is compared using GNSS RTK information.The experiment shows that the absolute pose error of the improved algorithm after GNSS fusion is reduced by an average of 43.9% compared to the improved algorithm without GNSS information fusion.更多还原