基于双目视觉的机器人目标定位技术研究

发布时间：2018-02-15 21:25

本文关键词： 双目视觉立体匹配 Hopfield神经网络目标定位三维重建　出处：《广东工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：机器视觉是机器人感知外界环境的一种重要手段,而深度信息是机器人利用机器视觉感知外界的一个重要信息。双目视觉是机器视觉领域的一个分支,也是一种获取深度信息的重要手段。双目视觉通过在不同角度观察同一目标,对产生的图像进行特征提取、立体匹配和视差计算,并进行重建三维场景来获取深度信息。双目视觉涉及的主要内容有:图像获取、相机标定、特征提取、立体匹配和三维重建。其中,立体匹配是双目视觉研究中的重点和难点。本文提出基于Hopfield神经网络的立体匹配算法,利用双目视觉,研究了机器人目标定位技术,并在此基础上进行了基于双目视觉的机器人目标定位实验。首先,本文对国内外基于双目视觉的机器人目标定位技术进行了深入研究,了解研究现状,总结分析了基于双目视觉的机器人目标定位技术中涉及的研究重点和难点;明确了本文的主要研究内容和研究工作。其次,研究了立体视觉基本原理,特别是双目视觉理论。在分析比较双目视觉系统原理的基础上,本文采用基于平行光轴的双目视觉系统作为机器人目标的定位的基本结构,并从几何角度分析了目标深度与视差之间的关系,奠定了双目视觉目标定位的理论基础。再次,重点研究了双目视觉目标定位涉及的关键技术,特别是坐标系统、相机模型、相机标定、双目相机标定与校正、三维重建等技术;另外,重点研究了基于张正友法的相机标定技术,并采用该方法对本文实验使用的相机进行标定。然后,深入研究了双目立体匹配,在分析研究了双目立体匹配基本原理、立体匹配约束、立体匹配难点和立体匹配算法的基础上,利用能量最小化方法,提出一种基于Hopfield神经网络的立体匹配算法;立体匹配中的极线约束、唯一性约束、平滑性约束和相似性约束引入到Hopfield神经网络能量函数中,不断更新神经元状态,从而最小化Hopfield神经网络能量函数,最终计算出视差图。最后,在理论研究的基础上,搭建了基于双目视觉的机器人目标定位实验平台;在平台上分别开展了相机标定、双目相机标定、基于Hopfield神经网络立体匹配和机器人目标定位等实验,并进行误差分析。在不断实验并分析误差,改进算法的基础上,完成了基于双目视觉的机器人目标定位,达到了令人满意的精度。
[Abstract]:Machine vision is an important means for robot to perceive the external environment, and depth information is an important information for robot to use machine vision to perceive external environment. Binocular vision is a branch of machine vision field. Binocular vision is also an important means of obtaining depth information. Binocular vision can extract features, stereo matching and parallax calculation of the generated image by observing the same object from different angles. The main contents of binocular vision are: image acquisition, camera calibration, feature extraction, stereo matching and 3D reconstruction. Stereo matching is an important and difficult point in binocular vision research. In this paper, a stereo matching algorithm based on Hopfield neural network is proposed. On this basis, the robot target localization experiment based on binocular vision is carried out. Firstly, this paper makes a deep research on the robot target location technology based on binocular vision at home and abroad, and understands the current situation of the research. This paper summarizes and analyzes the key points and difficulties involved in the robot target location technology based on binocular vision, clarifies the main research contents and research work in this paper. Secondly, the basic principle of stereo vision is studied. Especially the theory of binocular vision. On the basis of analyzing and comparing the principle of binocular vision system, this paper adopts the binocular vision system based on parallel optical axis as the basic structure of robot target localization. The relationship between target depth and parallax is analyzed from the angle of geometry, and the theoretical foundation of binocular visual target localization is established. Thirdly, the key technologies involved in binocular visual target localization, especially coordinate system and camera model, are studied emphatically. Camera calibration, binocular camera calibration and correction, 3D reconstruction, etc. In addition, the camera calibration technology based on Zhang Zhengyou method is studied, and the camera used in this experiment is calibrated by this method. The basic principle of binocular stereo matching, the constraint of stereo matching, the difficulty of stereo matching and the algorithm of stereo matching are analyzed, and the energy minimization method is used. A stereo matching algorithm based on Hopfield neural network is proposed, in which polar line constraint, uniqueness constraint, smoothness constraint and similarity constraint are introduced into the energy function of Hopfield neural network to update the neuron state. In order to minimize the energy function of Hopfield neural network, finally calculate the parallax map. Finally, based on the theoretical research, a robot target location experimental platform based on binocular vision is built, and the camera calibration is carried out on the platform. Binocular camera calibration, stereo matching based on Hopfield neural network and robot target positioning experiments, and error analysis. On the basis of continuous experiments and error analysis, improved algorithm, the robot target location based on binocular vision is completed. Satisfactory accuracy has been achieved.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】