基于立体视觉的深度探测系统设计与实现

发布时间：2018-06-13 17:16

本文选题：视差图 + Census变换　；参考：《北方民族大学》2017年硕士论文

【摘要】：基于立体视觉的深度探测在航空航天、辅助驾驶、虚拟现实等领域都有着广泛应用价值。对于深度探测系统的各个环节,国内外许多学者都做过大量的研究与实践,取得了突破性的进展。目前亟待解决的问题仍是如何同时满足其中核心算法立体匹配的准确度和实时性。本文搭建了一个简单的深度探测系统,其中所做的工作可概括为:根据定标结果双目校正,立体匹配获得视差图,三维点云重建获得深度信息并以此粗略估计地面高度。出于实时性的考虑,本文选用半全局立体匹配算法。并运用Middlebury大学目前最新的视差图评估系统对本文算法进行多组实验。本文主要研究成果可归纳为:1)在获取初步视差图中,本文采用相对灰度差与Census变换相结合的提取特征方式。并针对Census变换的权系数设计了多组实验,发现固定窗口半径和动态规划中惩罚代价影响Census变换权系数的最佳选取范围,以及在一定程度上增加Census变换权系数有助于检测轮廓变化剧烈或者分布密集的小物体。2)在获取最终视差图中,本文改进了基于RANSAC计算超像素视差平面的方法,提出先对初步视差值进行有控制的内插,填补视差图的部分空洞,再进行倾斜平面的建立的方式。改进之后,最终视差图物体边缘马赛克现象基本消失,轮廓变得更加自然、锐利。3)在地面高度测量中,本文采用对点云高度不同深度范围不同权值统计的方法,通过建立高度直方图,找高度值最小的点数波峰对应的高度氛围来确定地面高度。
[Abstract]:The depth detection based on stereo vision has been widely used in the fields of aeronautics and spaceflight, assisted driving, virtual reality and so on. Many scholars at home and abroad have done a lot of research and practice for each link of depth detection system, and have made breakthrough progress. At present, the urgent problem is how to satisfy the accuracy and real-time of stereo matching. In this paper, a simple depth detection system is built, in which the work can be summarized as follows: binocular correction based on calibration results, stereo matching to obtain parallax map, 3D point cloud reconstruction to obtain depth information and rough estimation of ground height. For the consideration of real-time, this paper chooses semi-global stereo matching algorithm. A new parallax evaluation system of Middlebury University is used to carry out many experiments on this algorithm. The main research results in this paper can be summarized as: (1) in obtaining the parallax map, we use the method of combining relative gray level difference and Census transform to extract features. Several experiments are designed for the weight coefficient of Census transform. It is found that fixed window radius and penalty cost in dynamic programming affect the optimal selection range of census transform weight coefficient. And to a certain extent, increasing the weight coefficient of Census transform is helpful to detect small objects with sharp contour change or dense distribution. In order to obtain the final parallax map, this paper improves the method of calculating superpixel parallax plane based on RANSAC. In this paper, a method is proposed to interpolate the initial parallax value, fill some holes in the parallax map, and then establish the tilting plane. After the improvement, the mosaic phenomenon on the edge of the parallax is basically disappeared, the contour becomes more natural, sharp. 3) in the measurement of ground height, this paper uses the statistical method of different weights in different depth range of point cloud height. The height of the ground is determined by setting up the height histogram and finding the height atmosphere corresponding to the point wave peak with the minimum height value.
【学位授予单位】：北方民族大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】