【作者】 孙玉娟；

【导师】 董军宇；

【作者基本信息】 中国海洋大学 ， 计算机应用技术， 2014， 博士

【摘要】 三维重建是计算机视觉领域的经典问题，其中基于单目相机的三维重建技术较其他技术更易被用户采用，采集数据时更方便。鉴于单目相机应用的普遍性和采集数据时的方便性，本文只研究单目视点下的三维表面重建技术。传统的基于单目视觉的三维表面重建存在较多的约束条件，特别是对光照环境的约束。这种约束条件会导致在进行数据采集时光源配置困难，配置的仪器设备价格较高。本文研究如何找到一种约束条件少，设备代价低，自然光或者复杂光环境下三维表面重建的解决方案；解决因遮挡，交互反射，阴影等因素引起的三维重建误差问题，准确且有效的恢复出物体的表面高度。本文主要工作和创新点包括：(1)综述了基于单目视觉的几种主流算法，这些算法是近几年基于单目视觉的具有代表性的算法，反映了近10年内基于单目视觉的最新技术，在单目视觉的三维重建中各有优势。本文基于这些算法对单目视点下的三维表面重建做了相应的改进。(2)介绍了几种常用的光照模型，并分析了不同光照模型的原理和特点。在各种光环境下拍摄了输入图像，基于不同的光照模型对输入图像进行了重绘和误差分析，分析结果表明球谐模型是对各种光环境较鲁棒的光照模型。(3)提出了基于参照物的非校准PMS三维表面重建算法，该方法将已知的参照物与目标物体放在同一场景中，并采集多幅输入图像。基于参照物可对输入图像的光照矩阵进行估计，可快速有效的估算出目标物体的三维形状。(4)将经典的PMS方法和光源参数估算方法合并，提出了一种快速估算人脸表面法向量的未校准PMS算法。通过在YaleB和BU3D数据库上的实验和分析，验证了人脸快速算法的有效性。(5)重新定义了基于耦合统计模型的框架，在此框架的基础上，可实现与训练库中光环境不同的单幅输入人脸的三维重建，对具有不同阴影效果的输入图像的重建结果鲁棒性较高。(6)提出基于拼接优化的单幅纹理三维重建算法。对岩石纹理进行了测试并与传统的SFS算法进行了对比，实验表明本文的算法对单幅输入图像的欠约束三维重建更有效。基于相机拍摄到的物体的图像，对目标物体进行三维重建，可将物体或者场景的三维形状准确的描述出来，去除由于环境的变化或者视角的偏差引起的对物体外观的理解错误，对于煤炭，钻井，勘探，考古等应用领域具有重要的应用前景。更多还原

【Abstract】 Three-dimention reconstruction is a classical problem in computer vision. Due tothe general and convenient application of the monocular camera, we only focus on3Dsurface reconstruction based on monocular vision technique in this paper.There are many constraints in the classical3D reconstruction based on monocularvision technology, especially for the constraints of lighting conditions, which willgenerate the difficulty in configuration of light sources and equipments. In this paper,we mainly focus on reducing these constraints, and making3D surface reconstructionbased on monocular vision more robust under the complex lighting conditions.Moreover the interactive reflection and shadow in3D reconstruction will also bediscussed. The main work and innovation points include:(1) The main algorithms based on monocular vision have been introduced, whichare the typical methods and have reflected the newest technology in the research ofmonocular vision. In this paper, we present our new method based on the ideas ofthese algorithms.(2) We discuss the commonly used lighting models and analyze theircharacteristics. By using these lighting models, the input images have been renderedunder various kinds of lighting conditions. Through analyzing the rendering error, themost robust lighting model has been selected to simulate the lighting condition of theinput image.(3) An uncalibrated PMS algorithm based on the reference object has beenproposed. Firstly, the target object and the reference object are put in the same scene,and the multiple images will be captured by making different lighting conditions.Then using the reference object, the lighting matrix will be estimated and the shape ofthe target object will be reconstructed quickly.(4) A fast uncalibrated PMS algorithm for estimating human surface normal hasbeen proposed by merging the classical PMS and the method of estimating lightingparameters. The effectiveness of this algorithm has been verified by the experimentsin the YaleB and BU3D databases.(5) A new framework of3D face reconstruction has been propsed based on a coupled statistical model. The3D shape can be estimated from a face image, whichhas the different lighting condition with the images in training set. The reconstructedresults are more accurate than the state of the art method.(6) An effective method has been presented for the same kind of objects. Thestitching and optimization has been used in the proposed method. An input rocktexture has been tested and compared with the method of SFS. The experimetnsverified the more effective of our proposed method than that of SFS.We can reconstruct the3D shape of the object from the captured images bymonocular camera. Then the intrinsic features of the object will be restored and not beaffected by the change of vision angle or the lighting conditons, which has animportant application prospection for coal, drilling, exploration and archaeology etc.更多还原