【作者】 程晓亮；

【导师】 赵杰；

【作者基本信息】 哈尔滨工业大学 ， 机械电子工程， 2006， 硕士

【摘要】 立体视觉技术是计算机视觉领域的经典课题,双目立体视觉由于具有仿人类双目的特点,成为立体视觉领域的热点。本文以移动机器人的双目立体视觉导航为背景,开展了立体标定、图像预处理、立体匹配及三维重构等技术的研究。在立体标定阶段,本文采用张正友的方法实现了单摄像机的标定并完成了图像畸变的修正,进而采用Fusiello的方法实现了双摄像机的标定并完成了立体图对的外极线校准。在图像的预处理阶段,本文对LOG滤波算法进行了改进,使用了一种四阶段的预处理方法,有效地消除了图像的噪声、平衡了立体图对间的亮度差异。在立体匹配阶段,本文以获得致密的视差场为目的,对区域匹配方法进行了详细的研究,比较了不同算子以及不同窗口对匹配性能的影响。在三维重构阶段,本文针对映射法的缺点,提出了一种对空间离散点集进行直接三角剖分的方法。该方法通过引入三角形法向量夹角、阈值距离以及最小内角等几个参数对三角形的生长进行约束,同时利用生长法,对空间中的离散点集进行逐步三角划分,直至所有三角形扩展完毕。实验结果表明,该算法使最优点的搜索范围大大降低,提高了三角形扩展的准确性,保证了三角网格的质量,可以准确高效的对空间中的离散点集进行直接三角剖分,具有很高的实用价值。同时,本文采用OpenGL作为可视化工具,使重构效果更加直观、形象。最后,我们在现有的双目视觉系统下,完成了包括立体图对的采集、标定、预处理、立体匹配及视差图的后处理、三维重构及可视化等全部环节在内的整合实验,验证了基于双目立体视觉进行三维重构的可行性,为完成移动机器人的视觉导航任务奠定了基础。更多还原

【Abstract】 Stereo Vision is the classical problem of Computer Vision Area. Binocular Stereo Vision becomes the famous research of Computer Vision Area, due to its eyes-similar characteristic. This paper is based on the navigation of mobile robot by Binocular Stereo Vision. It includes the research on calibration, image pretreatment, image matching and 3D reconstruction.In the stage of calibration, we use the method which was presented by Zhang Zhengyou, to calibrate the single camera and rectify the lens distortion. Furthermore, we use the method presented by Fusiello to do the double cameras calibration and epipolar rectification.In the stage of image pretreatment, we make an improvement of LOG arithmetic, and present a four-step image processing method. It can both eliminate the noise and balance the brightness difference, and get a high quality images finally.In the stage of stereo matching, in order to get dense disparity images, we do a lot of research on area matching method and compare the effect caused by different operators and windows.In the stage of 3D reconstruction, due to the disadvantage of mapping method, which is so-called "superiority" of triangulation in the 3D is not as good as in the 2D, in this paper, it presents a new approach that getting the triangulation of 3D scattered points directly. After the formation of the first triangle, use the growing method to extend the 3 edges of each triangle until all the triangles have been finished. It presents the angle of the two normal vectors formed by the new and the old triangles, the threshold distance and the principle of maximal-minimal angle in the paper to constrain the growing of the triangles, so as to reduce the searching area, improve the correctness and confirm the quality of the mesh. The method is proved to be highly exact and efficient. Also, we use OpenGL to make the model looks more vivid.更多还原