【作者】 李伟；

【导师】 马雷； 罗棕贵；

【作者基本信息】 燕山大学 ， 车辆工程（专业学位）， 2023， 硕士

【摘要】 自主地面车辆需要实时稳定地获取车辆的精确位姿,是实现自主导航和自主运动的必要条件。基于视觉惯性融合的位姿估计方法可以在GPS信号受限或完全丢失的环境中实现精确的位置估计。然而,单目视觉惯性位姿估计方法在地面车辆部分工况下存在尺度不可观的问题导致位姿轨迹退化。本文基于单目视觉惯性导航系统(Monocular Visual Inertial Navigation System,VINS-Mono)框架,对其进行改进,解决了视觉惯性融合应用于地面车辆中出现的问题,论文的主要研究工作如下:首先,基于Gazebo/ROS搭建了仿真车辆和室外仿真环境,旨在实现对自主地面车辆的运动仿真和传感器数据采集。此外,还搭建了基于阿克曼运动底盘的数据采集平台,该平台包含相机、轮速计、GPS和IMU传感器,可以采集模拟车辆在真实场景中的环境数据和运动数据。利用多个串口和多线程实现传感器数据高频通信,通过阿克曼车辆的运动学逆解来控制底盘运动,并在运动过程中收集传感器数据。接着,针对数据采集平台中相机、IMU和轮速传感器的测量模型和误差模型进行内参标定,并对比校准前后传感器数据。以相机为基准完成时空同步,将不同传感器的数据整合成一个完整的测量。然后,构建多传感器融合状态估计器,明确多传感器的测量约束和需要优化的状态变量。利用多传感器的互补特性完成视觉-IMU-轮速计联合初始化,为状态估计系统提供良好的初值。最后,基于仿真场景数据和实际场景数据进行测试,将本文算法与VINS-Mono进行对比。结果表明本文算法能有效提高车辆位姿估计系统的定位精度。更多还原

【Abstract】 Autonomous ground vehicles require real-time and stable acquisition of accurate vehicle posture,which is a necessary condition for achieving autonomous navigation and motion.A position and attitude estimation method based on visual inertial fusion has made significant progress in research and practical applications in the field of autonomous ground vehicle position and attitude estimation.However,the inertial pose estimation method based on monocular vision has a problem in that the scale is not observable under certain operating conditions of ground vehicles,leading to degraded pose trajectory.This paper aims to improve the framework of Monocular Visual Inertial Navigation System(VINS Mono)and address the challenges encountered in the application of visual inertial fusion to ground vehicles.The main contributions of this paper are as follows:Firstly,a simulation vehicle and outdoor simulation environment based on Gazebo/ROS for autonomous ground vehicles are built,which enables the collection of data for autonomous navigation.Additionally,a data acquisition platform based on the Ackermann Motion Chassis has been built,including cameras,wheel speedometers,GPS,and IMU sensors,to collect environmental and motion data of vehicles in real-world scenarios.Utilizing multiple serial ports and threads to achieve high-frequency communication of sensor data,controlling chassis motion through the inverse kinematics solution of the Ackermann vehicle,and collecting data during the motion process.Next,the camera,IMU,and wheel speed sensors in the data acquisition platform are calibrated using internal parameters based on measurement models and error models,and the sensor data are compared before and after calibration.Complete spatiotemporal synchronization based on the camera.Integrate the data of different sensors into a complete measurement.Then,a multi-sensor fusion state estimator is constructed to clarify the measurement constraints of multi-sensors and the state variables that need to be optimized.Utilizing the complementary characteristics of multiple sensors to complete the joint initialization of the vision IMU wheel speedometer,providing an accurate initial value for the state estimation system.Finally,the algorithm in this paper is tested and compared with VINS-Mono based on simulated and actual scene data.The experimental results demonstrate that the proposed algorithm is highly effective in enhancing the accuracy of vehicle positioning,thereby outperforming VINS-Mono.更多还原