港口装卸自动化中的机器人视觉导引系统的关键技术研究

【作者】 姜桂祥；

【导师】 杨忠根；

【作者基本信息】 上海海事大学 ， 通信与信息系统， 2003， 硕士

【摘要】 本文的立题来源于上海市教委的上海市高校科技发展基金项目《计算机视觉技术在港口集装箱装卸自动化中的应用》，它围绕用于集装箱装卸机器人进行视觉导引的几项关键技术展开研究。 本文所涉及的视觉导引系统主要由以下几个关键技术模块所组成：图像分割、边缘检测、优势点提取、摄像机标定、三维复原等。系统对由摄像机所摄取的图像进行一系列处理，从而获得目标的空间位置信息，提供机器人进行自主导引作业所需的距离、方向和摄像机相对目标的三维运动信息等数据。 图像分割和边缘检测是视觉技术的基础，也是决定整个系统性能优劣的关键。本文中，我们重点研究开发了结合熵理论的自适应平滑的边缘检测技术及基于模糊总熵准则的阈值分割技术，同时进行了大量的仿真实验，与现有的其他技术相比，效果明显改善；在优势点提取部分，我们开发了基于数据融合式的Fisher鉴别函数、基于最大巴氏距离以及基于小波技术的优势点提取技术，经过实验验证，它们均具有良好的应用性能；为了从目标视图精确测量它的三维信息，必须对摄像机进行标定，因此我们开发了一种无需任何精密设备或器材，仅需一个很易制作的平面标定板的标定技术，并在仿真过程中进行了检验；在本文的最后一章中，针对集装箱装卸的特殊应用环境，我们首先开发了几种单视图三维复原技术，包括：三点算法、非共面四点算法及共面四点算法；另外，为了适用于更为广泛的应用环境，我们又开发了用于双视图复原的降维八点算法，该算法不仅具有良好的去噪性能，能给出均方误差相当小的无偏估计，而且，与Hartley正则化技术相比，该算法运算速度更快，因此在实际应用中具有很大优势。这也是本文的重点研究内容之一。 本文主要研究了用于港口装卸自动化中的机器人视觉导引技术，由于其应用环境的特殊性，我们在研究过程中还特别注意了各种算法的鲁棒性及实时性，以适应风雨等恶劣天气状况及实际应用过程中对装卸速度的要求等。当然，文中的各种技术离实际应用还有一定距离，但考虑到目前国内的研究现状，我们的研究还是有一定的现实意义。更多还原

【Abstract】 In this paper, we develop some vital techniques of vision system guiding robot to autonomously complete container stevedore mission at port. The researched topic originates from the science and technology research fund project ’the application of computer vision to container-stevedore automation at ports’.The key modules of the visual guidance system are image segmentation, edge detection, breakpoint detection, camera calibration, 3D reconstruction, etc. When the images taken by camera have been processed by the system, the information of the spatial position and attitude of the interested object as well as the 3D motion speed of the camera relative to the object, which is necessary for guiding a manipulator to autonomously load and unload the containers, is extracted from them.Image segmentation and edge detection are the basic techniques of computer vision and their performances are vital for the result of the system. In the edge detection module, we study a new adaptive smooth algorithm, whose critical parameter is automatically selected in terms of entropy theory, and the threshold segmentation algorithm based on maximal total fuzzy entropy criterion. The experiments have demonstrated that they are superior to the traditional techniques. In the breakpoint detection module, we develop three techniques of breakpoint detection that are based on linear Fisher discrimination function, maximal Bhattacharyya-distance function, and wavelet analysis, respectively. In order to obtain the position information from the scene images accurately, the camera must be calibrated a priori. We develop a calibration algorithm, which doesn’t need any precise equipment but an easily made calibration board. And the calibration technology is found very effective during experiments. In the last module, according to the special environment of container stevedore, we develop several techniques of single-view 3D reconstruction, including the three-point, algorithm, the method based on four points lying at different planes or at the same plane. Furthermore, we develop the dimension-reduced EVD algorithm of two-view 3D reconstruction, which can efficiently suppress the effect from noise and give unbiased parameter estimation with small MSE. In addition, the algorithm is much simpler than and faster than the regularized algorithm developed by Hartley, so it is prevalent in practice. And it is one of the most important parts of this paper.Considering the vision guidance technique in this paper is developed for container-stevedore, the environment is very hostile and the stevedore must be real-time, we put the emphasis on the robustness and the real-time performance of each algorithm. Of course, there are still a lot of steps to be done before those techniques can be successfully applied in practice. But considering the present study situation in China, the work of the paper not only has attributed to the development of computer vision theory, but also has somepractical significance.更多还原