【作者】 李杨；

【导师】 陈剑；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2019， 硕士

【摘要】 随着计算机视觉及视觉传感器的快速发展,基于视觉的位姿估计及伺服控制越来越受学术界及产业界的关注及重视。可靠的视觉定位及伺服控制在无人驾驶、无人仓储、地质勘测等诸多领域有着重要的意义,是现代智能制造的亟需解决的关键技术之一。本文在考虑平面约束及相机内参不确定性的前提下,深入研究了基于多视图几何的视觉定位及伺服控制任务。在基于多视图几何模型的基础上,本文考虑了相机模型的内参不确定性问题,提出了一种新的位姿提取算法。同时结合车辆的平面运动约束,简化了位姿提取算法的求解复杂度。本文所提的方法,利用了多视图几何中帧间的关联特征信息,实现了两帧间的相对位姿估计。与传统的多视图几何方法相比,该方法同时适用于平面场景与非平面场景,且相机无需完全标定。为了解决相机内参不确定性的问题,本文引入了准归一化的欧式坐标系统,成功实现了像素平面与世界坐标系间的相对位姿参数化表示。进一步,针对该参数化方程,本文设计了求解器,完成了旋转信息量及平移信息量的独立求解。针对该算法,本文设计了随机采样一致性和数据归一化的优化算法,有效的提高了算法鲁棒性。本文通过仿真对比实验,将该算法与经典的多视图几何方法进行了对比。结果表明本文提出的方法在适用场景、求解精度上均有一定的优势,且在相机内参部分标定的情况下仍可奏效。本文在结合车辆非完整性约束的基础上,设计了移动机器人的运动学模型及闭环系统误差模型。同时,本文结合自适应时变控制器及前述所提算法,设计了视觉伺服控制任务的仿真实验及真实环境实验,完成了视觉伺服中的轨迹跟踪控制任务。仿真及真实环境实验结果表明本文提出的方法可以在内参部分标定情况下较好地完成视觉伺服常见的控制任务。综上,本文主要创新点及贡献包括:1.利用计算机视觉及视觉slam的相关技术,针对室内移动机器人,设计了一种基于单应性的视觉里程计,涵盖了图像增强、特征提取、特征匹配、光流跟踪、三维信息重构等环节。设计的视觉里程计通过真实环境下纯直线及弧线场景的实验,验证了其精度最高可达0.5mm/帧,完全满足移动机器人在工业环境中运作的性能要求。2.在传统多视图几何的基础上,原创性地提出了一种考虑内参不确定性且同时适用于共面与非共面场景的位姿提取算法。并提出了“准归一化”的欧式几何坐标系统,成功地解决了相机焦距参数未知下的位姿重构。针对该方法,提出了随机采样一致性和数据归一化的优化算法,并通过仿真实验,与经典的多视图几何算法进行了对比。3.结合移动机器人的非完整性约束,设计了移动机器人的运动学模型及闭环系统误差模型,并结合自适应时变控制器及前述所提算法,通过Matlab仿真及移动机器人Summit-XL的实车实验,实现了含内参不确定性情况下的视觉伺服轨迹跟踪任务。更多还原

【Abstract】 With the rapid development of the computer vision and vision sensors,vision-based pose estimation and visual servoing have gradually become one of the key technologies in the field of academic and industrial circles.Reliable visual positioning and control have important signifi-cance in many fields such as unmanned driving,unmanned storage,geological survey,etc.,and it is one of the key technologies that need to be solved in modern intelligent manufacturing.In this paper,considering the planar constraints and the uncertainty of intrinsic parameters,the visual positioning and control tasks are developed which are mainly based on multi-view geometry.Based on the multi-view geometry model,this paper considers the uncertainty of the intrin-sic parameters of the camera model and proposes a new pose estimation algorithm.Meanwhile,combined with the planar motion constraint of the vehicle,the complexity of the pose estimation algorithm is simplified.The method proposed in this paper makes use of the corresponding fea-tures between frames in multi-view to estimate the relative pose between two frames.Compared with the classical multi-view geometry based methods,this approach is suitable for both planar and non-planar scenes,and the camera can be partially calibrated.In order to solve the problem of intrinsic parameter uncertainty,this paper introduces a quasi-normalized European coordinate,which successfully realizes the parameterization of the relative pose relationship between the pixel coordinate and the world coordinate system.Furthermore,for the parametric equation,a solver is designed in this paper,and the solution of the rotational and translational information is devel-oped.Aiming at the proposed al,gorithm,this paper designs optimization algorithms including random sampling consistency and data normalization,which effectively improve the robustness of the proposed algorithm.This paper also compares the algorithm with the classical multi-view geometry based method through simulation and experimental results.The results show that the proposed method has advantages in applicable scenarios and solution accuracy,and it can still work with partially calibrated cameraBased on the nonholonomic constraint of vehicles,the kinematics model is designed.At the same time,this paper analyzes and compares different visual servo controllers,and realizes visual servoing tasks which applied with the proposed algorithm by simulation and experimental results.The visual servoing tasks mainly including trajectory tracking and target regulation.The simula-tion and experimental results int real environment show that the proposed method can successfully complete the tasks of visual servoing with partially calibrated camera.In summary,the main innovations and contributions of this paper are as follows:(1)Utilizing the techniques of computer vision and visual slam,a visual odometer was de-signed for mobile robot indoors which realize feature extraction,feature matching,optical flow tracking and pose extraction,and the accuracy is 0.5mm/frame.(2)Based on multi-view geometry,a pose estimation algorithm which considers the uncer-tainty of intrinsic parameters and both planar and non-planar scenes is proposed.And an opti-mization algorithm including random sampling consistency and data normalization is proposed for the pose estimation algorithm.The comparison with the classical multi-view geometry based methods are presented via simulation experiments.(3)Based on the nonholonomic constraint of the vehicle,the kinematics model was designed.And this paper successfully realizes the visual servoing tasks which applied with the proposed pose estimation algorithm by simulation and experimental results.更多还原