SIA OpenIR  > 机器人学研究室
地面机器人多模态融合 SLAM 方法研究
Alternative TitleMulti-modal Fusion SLAM for Ground Robot
苏贇
Department机器人学研究室
Thesis Advisor王志东
Keyword地面机器人 里程计 多模态融合 因子图优化 SLAM
Pages138页
Degree Discipline机械电子工程
Degree Name博士
2021-05-21
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract同步定位与建图(Simultaneous Localization and Mapping, SLAM)是机器人实现自主导航的前提与关键,近十年来,涌现出了许多优秀的基于视觉,激光雷达,以及多传感器融合的SLAM方法。但是现有的SLAM算法部署到地面机器人上时,由于机器人的运动退化、激光雷达退化、大尺度环境中视觉特征退化、全球导航卫星系统(Global Navigation Satellite System, GNSS)受到干扰、机器人轮子打滑等原因,使得现有的算法性能下降甚至失效。而且现有的融合机器人编码器的SLAM方法多在2D欧式空间中进行,对于存在斜坡、台阶、草坪、不平坦地形等3D环境则不能适用。因此本文针对地面机器人在复杂3D环境中的应用,对相机、激光雷达、惯性测量单元(Inertial Measurement Unit, IMU)、编码器、GNSS以及环境约束和先验知识等多模态测量融合的SLAM方法展开了研究。主要研究内容如下:(1) 在本文研究内容进行之前,还没有开源的针对地面机器人在室内室外各种复杂地形下的数据集。因此本文基于自主研发的机器人和感知平台,构建了地面机器人数据集GroundRobotDataset,具有16线激光雷达、双目相机、IMU、编码器、GNSS、实时动态差分(Real-time kinematic, RTK)等传感器的原始数据,包含了室内平坦环境、室内长走廊环境、室内以及室外斜坡、台阶、楼梯、减速带、不平坦地形、地下车库、室外开阔草坪、室外大尺度园区环境、动态目标等环境因素,还包含剧烈加减速及旋转运动、机器人轮子打滑、视觉失效、激光雷达退化、GNSS受到干扰等传感器数据异常情况。并且本文将此数据集开源。(2) 通过对机器人在不平坦的3D地形中的运动过程进行研究,提出了流形空间中的里程计增量模型,能够融合IMU与编码器测量计算机器人在3D空间中的位姿增量,适用于斜坡,楼梯等不平坦地形,将现有的基于编码器的位姿解算方法从2D欧式空间中扩展到3D流形空间中。并且基于所提出的里程计增量模型,提出了GR-SLAM算法,能够紧耦合视觉、IMU、里程计增量模型,在基于滑动窗口的因子图优化框架下进行机器人状态估计,提升了基于视觉的SLAM方法在地面机器人上的精度和鲁棒性。(3) 对基于激光雷达的SLAM方法进行了研究,并提出了GR-LOAM算法,紧耦合激光雷达,IMU,编码器的测量进行机器人的状态估计。利用各个传感器的局部估计结果进行互相监督,对机器人轮子打滑,激光雷达退化等异常传感器数据进行主动检测,实时调整各个因子的优化权重,实现了更加鲁棒的融合估计。本文充分利用机器人贴地运动约束,通过地面分割和地平面拟合,在优化中引入了局部地面约束,同时紧耦合全局重力约束,对机器人垂直方向的状态进一步进行优化。并且提出了一个高效的动态目标剔除算法,利用局部估计结果在不同时间维度上进行空间投影,根据匹配残差剔除位于动态目标的点和不稳定特征点,能够减小匹配误差,提高建图优化效率。(4) 根据视觉和激光SLAM的研究基础,对视觉与激光信息的互补融合进行了探索,并且提出了GR-Fusion算法,能够紧耦合激光雷达、相机、IMU、编码器、GNSS等多模态测量,并采用由局部到全局的优化策略进行机器人的状态估计。本文充分利用激光点云和拟合的地平面参数来提取视觉特征的深度,对视觉特征进行深度增强。并且根据景深与视角对视觉特征进行了分类,根据强约束方向构建为不同的残差因子,从不同方向对状态进行约束。局部优化时,选取高质量的视觉特征和激光特征构建约束因子,同时根据传感器退化情况自适应地调节不同模态的约束因子的数量和权重,并且紧耦合IMU、编码器、以及零速更新模型,保证局部优化的实时性和鲁棒性。全局优化时,利用滑动窗口同时将两帧细化后的激光点云与全局地图对齐,同时紧耦合局部增量约束与全局重力约束,对全局状态进行更加精确的低频优化。利用SLAM局部估计结果对GNSS置信度进行实时评估,通过融合GNSS测量与闭环约束对位姿图进行全局优化,消除机器人状态估计的累积误差。本文通过提出的GroundRobotDataset数据集以及实际的机器人平台,对所提出方法的性能进行了大量的实验验证。最后对本文的研究成果进行了总结,并且对未来的研究方向进行了展望。
Other AbstractSimultaneous localization and mapping (SLAM) is the prerequisite and key for robots to realize autonomous navigation. In the past ten years, many excellent SLAM methods based on vision, LiDAR, and multi-sensor fusion have emerged. However, when the existing SLAM algorithm is applied on the ground robot, due to the robot's motion degradation, LiDAR degradation, visual feature degradation in large-scale environments, Global Navigation Satellite System (GNSS) interference and robot wheels slipping, the performance of the existing algorithm decreases or even fails. In addition, the existing SLAM methods that incorporate robot encoders are mostly carried out in 2D Euclidean spaces, and are powerless for 3D environments with uneven terrain such as slopes, stairs, and lawns. Therefore, this dissertation focuses on the application of ground robots in complex 3D environments, and studies the SLAM method of multi-modal fusion. The main research contents are as follows: (1) Before the research in this dissertation, there is no open source dataset for SLAM on ground robots under various complex terrains indoors and outdoors. Therefore, based on the self-developed robot and perception platform, this paper proposes the ground robot dataset GroundRobotDataset, which has 16-line LiDAR, binocular camera, IMU, encoder, GNSS, Real-time kinematic (RTK) and other sensor raw data, including indoor flat environment, indoor Long corridor environment, indoor and outdoor slopes, steps, stairs, speed bumps, uneven terrain, underground garages, outdoor open lawns, outdoor large-scale park environments, dynamic targets and other environmental factors. It also includes sensor data abnormalities such as severe acceleration, deceleration and rotation, robot wheel slip, visual failure, LiDAR degradation, and GNSS interference. And this dissertation open source this dataset[131]. (2) By studying the motion process of the robot in uneven 3D terrain, an odometer incremental model on manifold is proposed, which can fuse IMU and encoder measurements to calculate pose increment of robot in 3D space, suitable for slopes, stairs and other uneven terrain, extending the existing encoder-based pose calculation method from 2D Euclidean space to 3D manifold space. And based on the proposed odometer incremental model, the GR-SLAM algorithm is proposed, which can tightly couple the vision, IMU, and odometer incremental model, and perform robot state estimation under the framework of factor graph optimization based on sliding window, which improves the the accuracy and robustness of the visual SLAM method on ground robot. (3) The SLAM method based on LiDAR is studied, and the GR-LOAM algorithm is proposed. The measurements of LiDAR, IMU, and encoder is tightly coupled to estimate the state of the robot. The local estimation results of each sensor are used to supervise each other, to actively detect abnormal sensor data such as robot wheel slipping and LiDAR degradation, and to adjust the optimization weight of each factor in real time to achieve a more robust fusion estimation. This paper makes full use of the robot's ground motion constraints, through ground segmentation and ground plane fitting, introduces local ground constraints in the optimization, and at the same time tightly couples the global gravity constraints to further optimize the vertical state of the robot. And an efficient dynamic target elimination algorithm is proposed, which uses local estimation results to perform spatial projections on different time dimensions, and eliminates points located on dynamic targets and unstable feature points based on matching residuals, which can reduce matching errors and improve mapping optimization effectiveness. (4) Based on the research foundation of vision and LiDAR SLAM, the complementary fusion of vision and LiDAR information is explored. The GR-Fusion algorithm is proposed, which can tightly couple multi-modal measurements, such as LiDAR, camera, IMU, encoder, GNSS, etc., and use local-to-global optimization strategies to estimate the state of the robot. This paper makes full use of the LiDAR point cloud and the fitted ground plane parameters to extract the depth of the visual features and enhance the depth of the visual features. Moreover, the visual features are classified according to the depth of field and the viewing angle, and different residual factors are constructed according to the direction of strong constraint, and the state is restricted from different directions. In local optimization, high-quality visual features and LiDAR features are selected to construct constraint factors according to the number of tracking and local roughness. At the same time, the number and weight of constraint factors of different modes are adaptively adjusted according to the degradation of the sensor. The IMU, encoder and zero-speed update model is also tightly coupled to ensure the real-time and robustness of local optimization. In the global optimization, a sliding window is used to align the refined LiDAR point cloud of the two frames with the global map at the same time, and the local incremental constraint and the global gravity constraint are tightly coupled to perform a more accurate low-frequency optimization of the global state. Finally, the GNSS measurements and loop constraints are integrated to optimize the global pose graph to eliminate accumulated errors. At the same time, the SLAM estimation results are used to evaluate the GNSS confidence in real time, and adjust its optimization weights adaptively. In this dissertation, a large number of experiments have been carried out to verify the performance of the proposed method through the proposed ground robot datasets and the actual robot platform. Finally, the research results of this dissertation are summarized, and the future research directions are prospected.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/28998
Collection机器人学研究室
Affiliation中国科学院沈阳自动化研究所
Recommended Citation
GB/T 7714
苏贇. 地面机器人多模态融合 SLAM 方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2021.
Files in This Item:
File Name/Size DocType Version Access License
地面机器人多模态融合 SLAM 方法研究(13606KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[苏贇]'s Articles
Baidu academic
Similar articles in Baidu academic
[苏贇]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[苏贇]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.