【作者】 张淑平；

【导师】 丁永生；

【作者基本信息】 东华大学 ， 模式识别与智能系统， 2010， 博士

【摘要】 位姿是反映并联机器人运动状态的重要参数,但由于并联机器人通常具有复杂的结构,其运动是三维空间的多自由度综合运动,且在无任何检测装置之前,并联机器人的精确空间位置是一个没有规律的未知数,所以如何同时获得多个自由度的高精度测量是并联机器人运动状态控制研究中亟待解决的重要难题。针对以上问题,本论文以机器视觉检测技术为手段,设计了基于视觉的并联机器人运动位姿检测系统,建立了并联机器人的特征提取和运动跟踪算法,而后分别提出了基于单目视觉和双目视觉的位姿估计方法。首先,论文系统地阐述了并联机器人的发展历史、国内外研究现状和主要的应用领域；分析了并联机器人位姿检测的方法及在并联机器人领域尚待解决的问题；描述了机器视觉系统的构成,理论的建立及机器视觉测量技术的研究与应用现状。其次,设计了基于整体信息处理机制的网络控制系统结构并针对并联机器人的位姿检测给出了视觉信息反馈框架及位姿描述方法。再次,建立并联机器人Haar-like特征集。而后利用该特征集训练简单分类器,并通过使用级联思想将各种简单的、独立的分类器结合起来构建了级联分类器,实现对并联机器人的运动跟踪。然后,讨论了基于单目视觉的三种并联机器人位姿检测方法：基于矩形不变量的位姿检测方法、基于点相关的位姿计算方法和基于免疫进化算法的位姿估计方法。其中方法一是利用并联机器人末端执行器上的矩形结构在投影中的两个仿射不变量来计算并联机器人的位置和姿态。方法二是以方法一所估计的并联机构的运动位姿信息作为迭代初值；以摄像机坐标系原点到末端执行器平面的距离作为变量,建立该变量与4个特征点的投影深度的关系模型；以末端执行器的刚体结构特性及4个特征点的共面特性建立了7个约束方程,最后由高斯迭代实现对变量的估计,获得并联机构的运动位姿信息。方法三是以待估计的并联机器人的6个位姿参数作为抗原；借鉴生物免疫系统中克隆变异和免疫记忆机理,通过免疫进化获得位姿参数的可行解。之后,构建了双目视觉位姿检测系统,提出了一种基于主动表观模型的并联机器人位姿检测方法。首先建立了分别适用于左右摄像机的两个AAM (Active Appearance Model)模型,并通过离线学习获得该模型参数；然后利用学习结果实现对输入图像中并联机器人的匹配和跟踪；最后根据投影空间的几何变换关系和匹配后的立体图像对,对并联机构末端执行器进行三维重建和位姿计算。最后,对全文的工作进行总结,并展望视觉检测领域中有待深入探索的几个研究方向。更多还原

【Abstract】 Pose is an important parameter reflecting the movement state of a parallel manipulator. However, the problem of measuring the full 6-DOF pose of a parallel manipulator with high accuracy is still urgent to be solved in the movement control of parallel manipulator due to the following three reasons:Firstly, the parallel manipulator usually has complex structure; Secondly, its movement is multi-degree-of-freedom comprehensive one in the three-dimensional space; Thirdly, when there is no any detection device, the precise space position of a parallel manipulator is an unknown quantity which does not obey any laws.To resolve these problems, a pose detection system based on the stereoscopic vision technology has been designed in the thesis for a moving parallel manipulator. Then the algorithms for feature extraction and tracking the moving parallel manipulator are established. Afterward, the pose estimation methods based on monocular vision and binocular vision are proposed to get high accurate movement pose parameters of the parallel manipulator.Firstly, the thesis illustrates the history, the current research and main application fields of the parallel manipulator systematically, analyzes the pose estimation methods and the problems to be solved in the parallel manipulator field, and describes the structure and theory of the machine vision system and the current situation of research and application of machine vision measurement technology.Then, we design a system structure diagram based on the integrated information processing mechanism. For the pose detection of a parallel manipulator, a vision information feedback framework and pose description methods are offered.Next, we extract the parallel manipulator’s different types of Haar-like features and set up a completed feature set. Then, we use this completed feature set to train simple classifiers, and sets up a cascaded classifier by different kinds of single and separate classifiers to track the parallel manipulator.Then, we discusses three methods of the parallel manipulator’s pose estimation based on the monocular vision:Vision-based pose identification of the parallel manipulator using two affine invariants of a parallelogram; Iterative pose estimation for parallel manipulator using points correspondences; Immune evolutionary algorithm to determine the position and rotation of parallel manipulator. In the first method, two affine invariants of a parallelogram on the parallel manipulator’s end-effector are utilized to determine the pose of the moving target relative to the camera. The second method is an iterative one. In the iterative process, results in the first method are used as initial values and the height of the vector from the camera’s focus point to the parallel manipulator’s end-effector is taken as the variable. Then, a model about this variable and the projection depths of each feature point is established, and an error matrix is also established through seven error functions which are produced by the depth estimation and the co-planarity of the four feature points. Finally, the Gaussian iteration is introduced to estimate the value of the variable, and to obtain the parallel mechanism motion pose information. In the third menthod, six pose parameters (three rotation parameters and three translation parameters) to be estimated are taken as antigens. Then several immune mechanisms (such as the clone and mutation mechanism, the immune remember mechanism) are borrowed to design an immune evolutionary algorithm. Through this evolution process, the best solution vector of six pose parameters can be obtained.Moreover, a pose measurement system based on binocular is built, and a method of calculating the parallel manipulator’s pose based on Active Appearance Model (AAM) is proposed. First, two independent AAM models are established:One for the right camera and the other for the left camera. Through an off-line training phase, the parameters of these two AAM models are obtained. Then the end-effector of parallel manipulator is tracked using the trained AAM models. At last, according to the geometry transform relations in projections space and the matched stereo image pairs, the 3D reconstruction and pose estimation of the parallel manipulator is realized.Finally, the work in this thesis is summarized, and several research directions which need to be explored further are listed.更多还原