【作者】 徐刚；

【导师】 刘彬； 李海滨；

【作者基本信息】 燕山大学 ， 光学工程， 2010， 博士

【摘要】 立体视觉技术是从二维图像中获取三维坐标信息的重要非接触测量手段,该技术通过对应点的匹配得到深度信息,实现物体轮廓或场景的三维重建。立体视觉技术在工业检测、逆向工程、航空测绘、医学成像和机器人视觉导航等很多领域都有广泛应用。基于立体视觉的稠密三维场景重建是机器人与环境信息交互的前提条件,特别是在非结构化的未知环境中,如何有效提高三维重建的计算效率对于移动机器人导航的实时性有着重要意义。本课题针对稠密三维重建中存在的计算复杂度问题进行深入研究,提出基于网格候选点的稠密匹配算法,并通过图割理论实现立体匹配的全局优化。具体工作包括:提出一种基于网格候选点的稠密三维重建方法,根据深度候选点在立体图像中对应点对之间的区域相关性获取稠密的深度信息。候选点的建立将对应点在二维空间的搜索转化为一维空间的匹配,解决了传统匹配算法中图像的极线校正和三维坐标求解步骤带来的计算复杂度。根据对应点在参考图像中的亚像素坐标计算节点三维坐标,实现候选点在深度方向的合理分布。针对稠密重建中用于匹配的候选点数量较大的问题,利用二维离散小波变换对图像进行塔式分解,并确定多分辨率下节点与各级分解图像的对应关系。通过对候选点由粗到精的逐级筛选,提高稠密匹配的计算效率;使用三目立体视觉系统代替双目,当互相关曲线的可信系数小于既定阈值,则根据另一组图像对的相似性测度对存在潜在歧义的少量候选点进行二次判决,消除匹配歧义;对于光照条件不理想或被测表面缺乏纹理的场景,采用结构光投影为场景主动添加纹理信息。依据对应点在图像坐标系的分布规律,研究投影图案对候选点匹配精度的影响,设计多区域锯齿波投影图案用于辅助测量,提高特定环境中三维重建的精度。研究三维网格节点与图割算法中网格图之间的关系,将网格节点直接用于构造网格图,并通过能量函数的最小化实现稠密重建的全局最优,解决低纹理区域和深度不连续区域匹配精度较差的问题。把筛选后剩余的节点构造成简化网格图,并依据候选点筛选过程中的归一化互相关值设定节点的能量函数,降低全局优化的计算复杂度。基于网格候选点的稠密三维重建方法降低了立体匹配的计算复杂度,有助于拓展自主移动机器人环境感知领域的应用。通过实验分析验证了三维场景稠密重建的效果。更多还原

【Abstract】 Stereo vision is one of the most important non-contact measurement techniques which acquire depth information from 2-D images. The 3-D contour of the object or scene can be reconstructed by corresponding point matching. The stereo vision technique was widely applied in many fields such as industrial inspection, reverse engineering, aerial mapping, medical imaging, robot navigation, etc. With stereo-vision-based dense reconstruction of the scene, the autonomous mobile robots have information exchange with the environment. How to improve the efficient of the 3D reconstruction is the key problem for the real-time requirement of the robot navigation, especially in the unstructured unknown environment. Aiming at the computational cost problem, the dense matching algorithm based on mesh candidate was proposed, and the graph cut was used for global optimization. The specific tasks include:A new 3D measurement method based on the mesh candidates was proposed. The depth can be recovered according to the region correlation of the correspondences. With depth candidates, the stereo correspondence was simplified from the 2D search problem to 1D search problem. The epipolar rectification of the images and the calculation of the 3D coordinate were no longer needed. The node coordinates in depth direction was determined by sup-pixel coordinates of the corresponding point in the reference image, which ensure the reasonable of the nodes distribution.According to the problem of the matching cost about large amount of the candidates in dense reconstruction, the pyramidal decomposition of the image was conducted using 2D discrete wavelet transform. The corresponding relationship between the nodes in different resolutions and decomposed images must be determined. By candidate screening from coarse to fine, the computational efficiency of dense matching can be improved. The binocular stereo vision system was represented by the trinocular one. If the confidence coefficient of the correlation curves is less than the preset threshold, the candidates of potential ambiguity can be reconsidered by another correspondence pair, by which the matching ambiguity was eliminated. In the circumstance with unsatisfactory light or the surface short of texture, the structured light was used to increase artificial texture. Based on the distribution of the corresponding points, the multi-block sawtooth pattern was designed for aided measurement. The precision of the 3D reconstruction improved by structured light in specific circumstance.By research on the relationship between mesh nodes and the graph, the graph can be established using nodes in world coordinate system. The global optimization of the dense 3D reconstruction would be fulfilled by minimization of the energy function, which solved the problem of the matching precision in low-texture region and discontinuous region. The simplified graph was constructed by screened nodes, and the energy function could be set by pre-calculated normalized cross correlation value. The algorithm complexity was decreased with a small quantity of the nodes. The accuracy and efficiency were verified by experimental result.The dense 3D reconstruction method based on mesh candidate simplified the computational complexity of the correspondence problem, expanding the application in the area of environmental awareness of the autonomous mobile robot. The performance of the dense reconstruction of the 3D scene was verified by experiments.更多还原