【作者】 杨帆；

【导师】 许男；

【作者基本信息】 吉林大学 ， 机械（专业学位）， 2024， 硕士

【摘要】 近年来随着汽车行业智能化、电动化的发展,自动驾驶辅助功能已经是各厂商和研究机构重要的研究方向。定位系统作为智能汽车的基础系统,其能否稳定地运行,提供精确的定位信息是汽车智能化中要解决的关键问题。本文的研究目的是在城市峡谷、隧道等GNSS(Global Navigation Satellite System)系统失效的情况下,采用单目相机、消费级IMU(Inertial Measuring Unit)和车辆的底盘数据作为定位系统输入,为智能车提供定位信息。当车辆处于退化运动状态时,视觉惯性定位会存在一定的不可观测性,所以引入了来自车辆底盘的数据作为额外的运动信息补充。本文首先对传统的车辆里程计进行了误差分析,通过相关参数的在线标定提高其定位精度,并将其得出的运动信息融入到视觉惯性SLAM(Simultaneous Localization And Mapping)中,提高了其初始化的稳定性和定位结果的精确性。本文的主要研究内容如下:(1)首先介绍了本文的研究背景,对智能车的定位技术进行了概述。对视觉SLAM中涉及的坐标系转换、位姿表示方法和位姿估计等内容进行了基础的理论介绍,并对本文所应用的具体传感器进行了原理介绍和参数标定,为后续的算法设计及实车实验提供了基础理论和实际参数的支持。(2)对传统车辆里程计的模型进行了介绍和误差分析,提出了动态的轮胎周长模型对其进行改进,并通过实验验证了相关参数改变对定位效果的影响。之后通过对相关参数建立非线性最小二乘问题,求解出待标定的参数优化值。对非线性最小二乘中涉及的传感器数据进行卡尔曼滤波,并对不同数据的噪声方差权重进行校正,再次对参数标定结果进行优化。(3)将改进后车辆里程计的数据融入到视觉惯性SLAM前端的尺度因子初始化过程中,减少车辆启动阶段IMU噪声产生的不利影响,提高了系统初始化的稳定性。在后端部分针对多传感器数据的融合,提出了视觉约束、惯性测量约束以及车辆的运动学约束,对重投影误差、IMU预积分和车辆预积分进行推导,将以上误差项整合在同一个最小二乘问题中进行优化。通过滑动窗口与边缘化操作,限制待优化状态量的数量,使系统在只对较新状态量进行优化的同时保留之前的相关观测约束,保证算法运行效率的同时提高输出定位结果的精度。(4)针对本文提出的算法进行验证与评估。搭建了实车数据采集平台,实地采集多种场景下的真实行驶数据。以高精度的千寻位置服务定位结果作为轨迹真值。本文算法与VINS-Fusion算法结果的对比,证明本文算法在初始化稳定性与车辆定位精度两方面上的提高,在有无回环,不同距离、环境的多个数据集中绝对轨迹误差都有29.59%以上的提升。更多还原

【Abstract】 In recent years,with the development of the automotive industry to intelligence and electrification,automatic driving assistance function has become an important research direction of various manufacturers and research institutions.As the basic system of intelligent vehicle,whether the positioning system can run stably and provide accurate positioning information is the key problem to be solved in automotive intelligence.The purpose of this paper is to use monocular camera,consumer IMU(Inertial Measuring Unit)and vehicle chassis data as inputs to the positioning system in the event of GNSS(Global Navigation Satellite System)system failure in urban canyons,tunnels,and other areas,in order to provide positioning information for intelligent vehicles.When vehicles are in degenerate motion state,there will be certain unobservability of visual inertial positioning.So,data from the vehicle chassis is introduced as supplementary motion information.In this paper,the error analysis of the traditional vehicle odometer is carried out,and the positioning accuracy is improved by online calibration of relevant parameters,and the motion information obtained is integrated into visual inertial SLAM(Simultaneous Localization And Mapping)to improve the stability of initialization and the accuracy of positioning results.The main research contents of this paper are as follows:(1)Firstly,the research background of this article was introduced.The basic theory of coordinate system transformation,pose representation and pose estimation in visual SLAM are introduced theoretically.The principle and parameter calibration of the specific sensor used in this paper are introduced.It provides the basic theory and practical parameter support for the subsequent algorithm design and real vehicle experiment.(2)An introduction and error analysis were conducted on the traditional vehicle odometer model,and a dynamic tire circumference model was proposed to improve it.The impact of relevant parameter changes on positioning performance was verified through experiments.Afterwards,a nonlinear least squares problem is established for the relevant parameters to solve for the optimized values of the parameters to be calibrated.The sensor data involved in nonlinear least squares is filtered by Kalman filter,the noise variance weights of different data are corrected,and the calibration results are optimized again.(3)Integrating improved vehicle odometer data into the scale factor initialization process of the visual inertia SLAM front-end reduces the adverse impact of IMU noise during vehicle startup and improves the stability of system initialization.In the back-end part,visual constraints,inertial measurement constraints,and vehicle kinematic constraints are proposed for the fusion of multi-sensor data.The reprojection error,IMU pre-integration,and vehicle pre-integration are derived,and integrated into the same least squares problem for optimization.By using sliding windows and marginalization operations,the number of state variables to be optimized is limited to ensure that the system only optimizes the newer state variables while retaining the previous relevant observation constraints,ensuring the efficiency of the algorithm and improving the accuracy of the output positioning results.(4)Verify and evaluate the algorithm proposed in this article.We have built a real vehicle data collection platform to collect real driving data in various scenarios.Using high-precision location service positioning results as trajectory truth.By comparing the results of the proposed algorithm with those of the VINS-Fusion algorithm,the results show that the proposed algorithm improves the initialization stability and vehicle positioning accuracy.In the presence or absence of loops,the absolute trajectory error of multiple datasets with different distances and environments can be improved over 29.59%.更多还原