【作者】 朱杰杰；

【导师】 潘志庚；

【作者基本信息】 浙江大学 ， 计算机应用技术， 2007， 博士

【摘要】 计算机视觉中的多视图几何技术构建了一套完整、成熟的从不同视点观察物体表面的理论。该理论囊括了许多几何知识，尤其是多幅视图之间存在的复杂而又美妙的几何关系已成为计算机理解客观世界的重要途径之一。本论文针对该技术在增强现实中的需求开展研究工作，取得的研究成果包括：(1)提出了基于数据不确定性的约束求解模型：数据是客观实体信息的载体，由于采样的不精确性，使得它包含了不确定性。在大多数视觉几何约束求解方法中，都忽略或简化了它对精度的影响。本文从数据不确定性出发，分析出图像不确定性的因素，提出了一个基于数据不确定性的约束求解模型(DUCRM)。该模型推导出4种不同层次的二维数据不确定性，并定义了图像不确定性的表达及其几何含义；通过迭代求解时不确定性的非线性扩散结果，DUCRM对观测向量进行了逐步求精，并提高了求解精度。(2)提出了改进的平面单应约束与对极几何约束的求解算法：针对低分辨率图像的兴趣点检测难、匹配精度低的问题，本文采用基于Harris角点检测与高阶曲线内部约束的混合算子来提高检测与匹配的精度，并结合DUCRM模型提高平面单应约束的求解精度。对极几何约束的线性求解方法对噪声敏感、精度较低；而非线性方法虽然精度高，但计算量大；另外，通用DUCRM模型求解时的奇异矩阵容易导致数值不稳定。针对这些问题，本文提出了一种常数项消除的方法避免了奇异矩阵，该方法根据观测向量的权值中心推导出待求参数矢量的最后一个元素与前8个元素的关系，从而分离出最后一个元素。实验结果表明，结合扩展DUCRM模型，该方法能较大地提高求解精度。(3)提出了一种视频增强现实中虚实物体遮挡处理的加速算法：虚实物体之间正确、快速、鲁棒的遮挡处理一直是增强现实的难点之一。基于深度图求解方法的关键是如何快速地求出真实物体的深度，本文针对普通PC机提出了一种深度近似的快速方法：先用全局与局部高斯核对前景物体进行提取；接着用提取的物体轮廓线上的点进行立体匹配，立体匹配采用对极几何约束缩小搜索范围；再用三角法将匹配结果反投到空间中，并求出深度均值作为前景物体的深度值；最后根据深度图以及物体的重叠区域进行遮挡处理。该方法具有速度快、精度高的优点。(4)提出了一种基于反光球的快速真实光照检测与阴影绘制方法：针对双向反射分布函数、基于图像的重光照等方法恢复真实光源计算量大，且必须已知部分场景几何的问题，本文提出一种基于反光球的快速方法：假设真实光源离散地分布在半空间中，根据光源在反光球上的分布情况，可以快速地估计出真实光源的位置和方向。本文将该方法应用于开发的视频增强现实原型系统中，实现了虚拟物体的快速阴影绘制。本文最后开发了一个增强型海底世界交互系统。它是一个可以容纳多个体验者共同参与，并能与各种虚拟海洋生物进行交互的系统。它综合实现了上面提出的几个算法，并实现了主体人与虚拟化身之间的和谐交互。该原型系统的开发探索了类似系统在科技馆中展出的可能性，具有潜在的商业应用价值。更多还原

【Abstract】 Multiple View Geometry in Computer Vision has build an all-around and sophisticmapping model between shapes from projective objects and camera’s parameters. Thismodel is composed of many geometry knowledge, particularly the complex and grace-ful geometric relationship among several views. It has also become one of the mostimportant fundamental theories of understanding real world for computers. This dis-sertation studies the application of Multiple View Geometry in Augmented Reality, andmain contributions are:(1) A new vision parameter estimation method using Data Uncertainty-basedConstrain Resolver Model (DUCRM): Data describes the characteristics of real objects.It also carries the uncertainty which is not wanted. In most of current estimation meth-ods, data uncertainty is overlooked and simplified. The accuracy of estimated param-eters thus is affected. This dissertation defines the analytic representation of data un-certainty on image and provides the algebraic equations for uncertainty propagation innonlinear functions, which are all based on data uncertainty theory. It provides a four-layer approach and a revising scheme in minimization iterative process. The accuracyof estimated parameters are thus improved.(2) Improved planar homography estimation algorithm using mixed operatorand improved epipolar geometry estimation algorithm based on constant elimina-tion: Calculating planar homography between two low resolution images is always adifficult problem because of larger errors in comer detection and matching. We pro-vides a mixed detection algorithm using Harris comer and curve functions to achievebetter results. And with DUCRM, we get a better result compared with other methods.In addition, Epipolar Geometry is also an important vision parameter for perspectiveprojective applications. Traditional linear methods is simple but subject to noise per-turbance. And nonlinear methods give better result by a global searching, which iscomputing expensive. Experiment shows that simply using DUCRM leads to singularproblems which cause numerical calculation unstable. We provide an approach by elim-inating the constant column in measurement function. Results show that the singularmatrix is avoided and computing speed is accelerated.(3) An accelerated algorithm of occlusion handling between virtual and real ob-jects: Correct, fast and robust occlusion handling between virtual and real objects isalways a critical problem in interactive AR. How to accelerate the computing speed is the key of using depth map based methods. We provides an approximation methodwhich can be implemented on PCs. First, a global and local Gaussian kernel function isused to extract foreground objects. And then, stereo matching is conducted only usingpoints on objects’ contour. This process is also accelerated by Epipolar Geometry. Fore-ground object’s depth is thus estimated using triangle method and occlusion is detectedby the overlapped region. Compared with other methods, ours is fast and gives betterresults. However, since one object has only one depth value, this method can not beapplied to deformable objects.(4) A fast shadow rendering algorithm in dynamic scene based on reflective lightball: Although BRDF and Image-based Relighting methods have good models to re-cover light sources, they can not done in real time as for a dynamic scene. We providesa discrete light sample method to approximate the light source in real scene, which us-ing a light reflection ball. This method is applied in a Video-based Augmented Reality(VAR) system simulated by PCs without GPU acceleration.Based on the estimation model and algorithms described above, we also developedan interactive Video-based Augmented Reality prototype. It is a mixed environmentwhich provides tools to enable observers walk through and interact with virtual seaanimals. In this prototype, algorithms are implemented and tested. And interactivescheme between observers and virtual avatars is also designed and implemented. Thisprototype helps us explore the feasibility of creating and exhibiting such systems inScience Centers.更多还原