【作者】 刘芳；

【导师】 秦勇；

【作者基本信息】 哈尔滨理工大学 ， 仪器科学与技术， 2020， 硕士

【摘要】 实时定位与地图创建(SLAM)系统在移动机器人领域应用广泛。单目视觉SLAM系统因其计算量小,结构简单等优势成为当前的研究热点,其缺点是在场景纹理弱或运动过快时容易丢失信息,且单目相机无法获得环境的绝对尺度。而惯导传感器(IMU)可估计自身角速度和加速度从而获得绝对尺度信息,且在缺乏相机信息的环境中可提供位置信息,同时相机信息可以修正惯导的累积漂移。基于此,本文研究了一种单目视觉和惯导数据融合的SLAM方法,可以提高SLAM系统的定位精度和鲁棒性。本文研究了相机姿态估计以及相机和IMU的标定方法。首先,采用KLT跟踪算法对ORB算法提取到的特征点进行匹配,利用MSAC算法对匹配过程中的误匹配进行了剔除,克服了目前常用的去除误匹配算法(RANSAC)存在的阈值选取过于敏感的问题。其次,针对目前求解相机姿态估计算法计算量较大的问题,引入了改进的DLT算法对相机的姿态进行估计,与当前主流的EPn P算法相比,有效地提高了姿态估计的运算速度。最后,研究了利用Kalibr工具标定相机和IMU的方法。在相机姿态估计的基础上,研究了视觉和惯导的联合初始化及优化。首先,采用纯视觉的SFM算法对相机信息进行初始化,获得SLAM系统运行所需初值。其次,利用旋转矩阵进行相机与惯导数据的联合初始化,将相机坐标系和IMU坐标系都变换到世界坐标系下。最后,采用相机与惯导数据紧耦合的方式构建优化方程,减少了传感器因采样频率不同引起的局部漂移问题。针对系统全局漂移问题,通过回环检测和全局姿态优化方案保证系统整体的一致性。通过搭建实验平台对基于单目视觉和惯导数据融合的SLAM方法进行验证。首先对相机和IMU进行标定实验,确定了相机的内参和外参矩阵,并且得到其重投影误差不超过一个像素点。然后利用Eu Roc数据集和实际的环境进行定位和建图,验证了本文的SLAM方法精度高,鲁棒性好。更多还原

【Abstract】 The Simultaneous Localization And Mapping(SLAM)system is widely used in the field of mobile robots.Because of its small computation and simple structure,monocular vision slam system has become a research hotspot.But,its disadvantages is that it is easy to lose information when the scene texture is weak or the motion is too fast,and the monocular camera cannot obtain the absolute scale of the environment,The Inertial Measurement Unit(IMU)can estimate its own angular velocity and acceleration to obtain absolute scale information,and can provide position information in the environment of lack of camera information.At the same time,the camera information can correct the accumulated drift of IMU.Based on this,this dissertation studies a slam method of monocular vision and INS data fusion,which can improve the positioning accuracy and robustness of slam system.In this dissertation,camera pose estimation and camera and IMU calibration methods are studied.Firstly,KLT tracking algorithm is used to match the feature points extracted by ORB algorithm.The MSAC algorithm is used to eliminate the mismatches in the matching process,which overcomes the problem that the threshold selection of RANSAC algorithm is too sensitive.Secondly,in order to solve the problem of large amount of computation in the current camera pose estimation algorithm,the improved DLT algorithm is introduced to estimate the camera’s pose.Compared with the current mainstream EPNP algorithm,the algorithm speed of pose estimation is effectively improved.Finally,the method of calibrating camera and IMU with kalibr tool is studied.Based on the camera pose estimation,the joint initialization and optimization of vision and IMU are studied.Firstly,the camera information is initialized by the vision-only SFM algorithm to obtain the initial value needed for the operation of the slam system.Secondly,the camera coordinate system and IMU coordinate system are transformed to the world coordinate system through the joint initialization of the camera and INS data by using the rotation matrix.Finally,the optimization equation is constructed by the way of tight coupling between camera and INS data,which reduces the local drift caused by different sampling frequency.Aiming at the global drift of the system,a scheme of loop detection and global pose optimization are used to ensure the overall consistency of the system.The slam method based on monocular vision and INS data fusion is validated by building an experimental platform.Firstly,the calibration experiment of camera and IMU are conducted,and the internal and external parameter matrix of the camera is determined,and the re-projection error is less than one pixel.Then,using the data set of Eu Roc and the actual environment to locate and map,the high precision and the good robustness of the slam method in this dissertation are verified.更多还原