【作者】 王越；

【导师】 王向军；

【作者基本信息】 天津大学 ， 仪器科学与技术， 2020， 博士

【摘要】 相较于成熟的雷达法、激光扫描法,基于计算机视觉的空间坐标测量方法具有非接触、全场式、高精度、低成本等优势,而针对大尺度视觉空间的空间坐标测量,静态立体视觉系统的基线距离固定,视场范围受限。为扩展视觉测量范围与提高测量精度,本文提出一种基于旋转、运动、非变焦摄像机的动态立体视觉空间坐标测量方法。本文针对动态立体视觉系统的误差源、摄像机的初始参数校准、摄像机的自动对中、摄像机在旋转或移动后动态立体视觉系统的外部参数校准与空间坐标测量等关键问题展开研究。本文主要的研究内容如下:(1)针对动态立体视觉系统的空间坐标测量,本文提出一种简化的基于差分GPS的动态立体视觉空间坐标测量模型,并将系统的误差源归结为静态误差与动态误差,着重研究各误差因素的权重以及误差因素与视场中心邻域空间坐标重建精度的关联。针对动态立体视觉系统的初始参数校准,本文提出一种六点法,即采用六个已知三维信息的控制点预先估计各个摄像机的焦距与初始姿态角,实验验证了该方法的可行性与鲁棒性,实验数据表明焦距与姿态角的标准差分别不超过0.05mm与0.019°,并且测量距离约为650m时,空间坐标的均方根误差不超过0.4m。(2)本文采用基于图像配准的目标匹配方法实现动态立体视觉系统中摄像机的自动对中,即采用不同场景图像的对应特征点计算图像间的单应矩阵并估计目标在其他场景图像中的位置从而控制摄像机旋转使得目标位于各个摄像机的视场中心。针对图像间的特征点匹配,本文结合SURF算法中利用Hessian矩阵提取特征点的方法与ORB算法中利用steered Brief算法生成描述子的方法,并取代传统的RANSAC方法,提出一种基于摄像机几何约束条件的误匹配剔除方法。相较于SURF、ORB算法,Mikolajczyk数据集仿真实验表明本文方法能够剔除相似程度较大的误匹配点对,并且耗时最少。针对目标位置估计,本文取代全局单应矩阵并采用基于Moving DLT的局部单应矩阵来描述图像间的映射关系,提高了目标的定位精度。(3)为实现摄像机旋转后动态立体视觉系统对空间坐标的实时测量,本文提出一种利用自然场景中的一组相交直线进行旋转摄像机外部参数校准的在线自标定方法。该方法基于摄像机旋转前后的内部图像单应性,仅利用相交直线的单个交点与摄像机旋转角度的初始值迭代估计摄像机的外部参数,并根据立体视觉的基本矩阵特性,采用Nelder-Mead无约束优化算法对外部参数进行修正。实验数据表明该方法估计的旋转矩阵欧拉角与平移向量相对于参考值的绝对误差平均值分别不超过0.054°与7.2mm,并且在测量距离约为200m时,摄像机旋转后的空间坐标均方根误差低于0.4m。(4)为实现摄像机移动后动态立体视觉系统对空间坐标的快速测量,本文提出一种仅利用单个已知三维信息的控制点进行运动摄像机外部参数校准的线性算法。该线性算法与仅利用单个同等控制点的迭代算法具有同等精度,但运行时间缩减了94.5%。实验数据表明该方法估计的旋转矩阵欧拉角与平移向量相对于参考值的绝对误差平均值分别不超过0.01°与2.5mm,并且在测量距离约为100m时,摄像机移动后的空间坐标均方根误差低于0.28m。更多还原

【Abstract】 Compared with mature radar method and laser scanning method,the method of spatial coordinates measurement based on computer vision has the advantages of non-contact,full-field,high accuracy,and low cost.For the spatial coordinates measurement in large scale visual space,the baseline distance of the static stereo vision system is constant,and the field of view(FOV)is limited.In order to extend the range of visual measurement and improve the measurement accuracy,this paper presents a method for measuring spatial coordinates of dynamic stereo vision(DSV)based on rotating,moving and non-zoom cameras.In this paper,the error sources of the DSV system,the initial parameters calibration of the cameras,the automatic alignment of the cameras,the extrinsic parameters calibration and the spatial coordinates measurement of the DSV system after the cameras rotate or move are studied.The main research contents of this paper are as follows:(1)For spatial coordinates measurement of DSV system,this paper proposes a simplified spatial coordinates measurement model of DSV based on differential GPS,and the error sources of the system are summarized as static error and dynamic error,and the weight of each error factor and the correlation between the error factors and the reconstruction accuracy of the spatial coordinates of the central neighborhood of the FOV are emphatically studied.In view of the initial parameters calibration of the DSV system,this paper proposes one six-point method,which six control points with known three-dimensional(3D)information are used to estimate the focal length and initial attitude angles of each camera in advance.The experiment verifies the feasibility and robustness of this method,experimental data reflects that the standard deviations of the focal length and attitude angles are respectively not more than 0.05 millimeter and 0.019°,and the root mean square error(RMSE)of spatial coordinate is less than 0.4 meter when the measurement distance is about 650 meter.(2)In this paper,an object matching method based on image registration is adopted to realize the automatic alignment of the camera in DSV system,which the corresponding feature points between different scene images are used to calculate the homography matrix between images and estimate the position of the target in other scene images so as to control the rotation of the camera so that the target is located in the center of each camera’s FOV.In view of the feature points matching between images,this paper combines the method of extracting feature points using the Hessian matrix in SURF algorithm and the method of generating descriptors using the steered Brief algorithm in ORB algorithm,and proposes one method of eliminating mismatches based on camera geometric constraints instead of traditional RANSAC method.Compared with the SURF and ORB algorithms,the simulation experiment of Mikolajczyk dataset reflects that the method in this paper can eliminate the false matching point pairs with a large degree of similarity and takes the least time.Aiming at the estimation of the target position,this paper replaces the global homography matrix and uses the local homography matrix based on Moving DLT to describe the mapping between images,which improves the positioning accuracy of the target.(3)In order to achieve real-time measurement of spatial coordinates of the DSV system after the cameras rotate,an online self-calibration method is proposed to calibrate the extrinsic parameters of rotating cameras by using a set of intersecting lines in the natural scene in this paper.This method is based on the inter-image homography before and after the rotation,the camera’s extrinsic parameters are estimated iteratively by only using the single intersection of the intersecting straight lines and the initial values of the camera’s rotation angles,and the Nelder-Mead unconstrained optimization algorithm is used to modify the extrinsic parameters according to the basic matrix characteristics of stereo vision.Experimental data reflects that the average of absolute error of the Euler angles of rotation matrix and the translation vector estimated by this method with respect to the reference values are respectively less than 0.054° and 7.2 millimeter,and the RMSE of spatial coordinate after rotation is less than 0.4 meter when the measurement distance is about 200 meter.(4)In order to realize the rapid measurement of spatial coordinates of the DSV system after the cameras move,this paper proposes a linear method for calibrating extrinsic parameters of a moving camera using only a single control point with known3 D information.This linear algorithm has the same accuracy with the iterative algorithm using only the same single control point,but the running time is reduced by94.5%.Experimental data reflects that the average of absolute error of the Euler angles of rotation matrix and the translation vector estimated by this method with respect to the reference values are respectively less than 0.01° and 2.5 millimeter,and the RMSE of spatial coordinate after moving is less than 0.28 meter when the measurement distance is about 100 meter.更多还原