【作者】 王杰；

【导师】 都思丹；

【作者基本信息】 南京大学 ， 电子与通信工程（专业学位）， 2021， 硕士

【摘要】 立体视觉是计算机视觉领域被广泛研究的课题之一,其在机器人避障、三维重建以及自动驾驶等领域有十分重要的应用。作为立体视觉的核心问题,立体匹配算法在很大程度上决定了整个系统的精度与耗时,而巨大的计算量、遮挡、弱纹理以及光影问题又对立体匹配提出了巨大的挑战,传统的双目立体匹配算法难以克服以上问题。在本文中,我们基于多基线三目相机模型,对于立体匹配算法进行优化改进以解决如上问题,并着手搭建了三目立体视觉系统使算法落地。具体成果如下:(1)基于平行多基线三目相机模型,提出了三目动态视差范围优化方案用于加快立体匹配算法的运行速度。该方案将窄基线相机生成的视差图作为先验知识,为宽基线相机组的图像的每个像素点计算出缩小后的动态视差搜索范围,后续宽基线的立体匹配将在这个缩小后视差范围内计算,从而可以实现在能计算出视差真值的情况下极大减少算法计算量的目的。基于此方案优化改进后的三目立体匹配算法可以在保持视差图精度基本无损的情况下,大大减少计算时间,显著提高算法效率。(2)为提高视差图精度,应对遮挡、弱纹理以及光影区域带来的误匹配问题,我们提出了基于非共线多基线三目相机模型的三目视差置信度估计的优化方案。该方案目的在于建立一套评估视差值可信程度的测量方法,借助于三目相机模型横纵双基线相机组的特点,使其生成的视差图可以优势互补,从而可以显著提高视差图的精度。基于此方案优化改进后的算法,可以极大改善遮挡、弱纹理、光影等区域的误匹配情况,同时通过并行计算又可保证其计算速度。(3)实践了三目立体视觉系统的落地搭建过程,研究设计了系统方案,探究了三目相机标定的原理并完成了系统的立体标定。更多还原

【Abstract】 Stereo vision is one of the widely studied subjects in computer vision areas,and it plays an important role in many applications,such as robot obstacle avoidance,3D reconstruction and automatic driving.As the core problem of stereo vision,stereo matching algorithm to a large extent determines the accuracy and time consumption of the whole system,while the huge computational complexity,occlusion,weak texture,light and shadow problems make stereo matching a great challenge.The traditional binocular stereo matching algorithm is difficult to overcome these problems.In this paper,we optimize the stereo matching algorithm based on the multi-baseline trinocular camera model to solve the above problems,and build a trinocular stereo vision system to implement the algorithm.The specific results are as follows:(1)Based on the horizontal multi-baseline trinocular camera model,an optimization scheme named trinocular dynamic disparity range is proposed to speed up the stereo matching algorithm,In this scheme,the disparity map generated by the narrow-baseline camera is taken as the prior knowledge,and the reduced dynamic disparity search range is calculated for each pixel of the image of the wide-baseline camera,which will be used for wide-baseline stereo matching and can greatly reduce the computational complexity of the algorithm when the ground truth is included.The improved algorithm based on this scheme can greatly reduce the time consumption and significantly improve the algorithm efficiency while the accuracy of the disparity estimation is basically unchanged.(2)In order to improve the accuracy of disparity estimation and deal with the problem of mismatching caused by occlusion,weak texture and light and shadow regions,we proposed an optimization scheme named trinocular disparity confidence measure based on the non-collinear multi-baseline trinocular camera model.The purpose of the scheme is to establish a set of measurement methods to evaluate the reliability of the disparity value.By virtue of the characteristics of the horizontal and vertical binocular cameras of the trinocular camera model,the disparity map generated by the model can be complementary to each other and the accuracy of the disparity estimation can be significantly improved.The optimized algorithm based on this scheme can greatly improve the mismatching of occlusion,weak texture,light and shadow problems,and at the same time,the calculation speed can be ensured by parallel computation.(3)We complete the construction process of the trinocular stereo vision system,research and design the system scheme,explore the principle of the trinocular camera calibration and complete the stereo calibration of the system.更多还原