【作者】 周佳立；

【导师】 张树有；

【作者基本信息】 浙江大学 ， 机械设计及理论， 2009， 博士

【摘要】 本文以对物体表面进行三维曲面重构为目的,构建不同的双目被动立体视觉系统,对视觉系统的立体标定、图像匹配、密集点云重建、细分曲面拟合等方面展开了深入的研究,并对细分法的快速收敛性判据的计算方面做了一些有益的探索工作。最后给出了双目立体视觉系统在三维人脸识别中的实例应用。本文的主要工作如下:提出基于复小波的相位相关算法对图像对进行密集匹配。考虑顺序匹配约束、连续性约束和相关性约束等条件,通过函数波峰拟和与亚像素位移因子迭代技术,对图像进行多分辨小区域匹配,获得对应点的亚像素级实时配准效果。该方法较好地克服了由传统傅里叶变换周期性引起的边界跳跃等问题,又较Gabor变换具有从粗到细的多分辨匹配搜索能力,对光线变化等具有较强的鲁棒性,可密集重建物体表面三维点云信息。最后运用细分曲面拟合技术,对物体表面进行层次局部加细的曲面重建。构建一种基线自适应的双目被动立体视觉系统,可根据被测物体的采样距离与所需精度等参数通过单片机自适应地驱动基线长度,由于调整基线和拍摄抖动会引起系统立体标定参数的改变,需引入新的半自标定技术,该技术可实时精确获取新的标定结果。另外,为解决传统双相机精确同步采样困难的问题,构建一种基于单相机的双目被动立体视觉系统,并为此提出了与之相应的标定算法,解决了非针孔模型下的相机自动标定问题。对细分法的收敛性分析也做了相关的工作。p-范数联合谱半径是由复方阵的有界集定义的,可用来判别细分法的收敛性。这里研究了整数值的p-范数联合谱半径,给出一些基本公式,并对∞-范数下的Berger-Wang关系式给出一个简单的证明。另外,对细分随机矩阵特征值进行了估计。最后,作为对本文研究内容的实例应用,提出了一种新的三维人脸识别方法。该方法采用基于单个相机的双目立体视觉系统对人脸进行采样,根据人脸对称性假设,运用补洞与纠错技术进行自动点云优化,继而采用简化的CANDIDE-3模型作为细分初始控制网格,局部加细地进行细分曲面分层次拟合操作,采用测地线映射技术对不同表情进行归一化,并分别建立人脸数据库。运用从粗到细的策略进行三维人脸的比对与识别,该识别方法也同时支持通过图像序列反求的三维人脸信息。实验结果表明,采用新的单相机立体视觉系统在提高重建精度的同时,很大程度上避免了由于双相机拍摄不同步引起的重建鲁棒性降低问题。而采用细分曲面作为存储结构,在节约空间的前提下,为分层次比对筛选提供了理论支持。该系统成本较低,适合在许多领域推广应用。更多还原

【Abstract】 In this paper, a different passive binocular stereo vision system to get 3D surface reconstruction is presented. An in-depth research is done in vision system calibration, image matching, dense point cloud reconstruction, subdivision surface fitting and so on. Moreover, there is also some useful work on rapid computation of convergence criterion of subdivision rules. Finally, some examples are given by applying the binocular stereo vision system to face recognition.The main research achievements are as follows in detail:Based on complex wavelet, a phase correlation algorithm is put forward to match images intensively. Considering order matching constraint, continuity constraint and correlation constraint conditions, multiresolution image regional matching is done through peak fitting and iteration of sub-pixel displacement factor, which leads to sub-pixel real-time matching results for corresponding points. This method makes it feasible to overcome boundary jumping and other problems caused by the periodicity of traditional Fourier transform. Meanwhile, it has coarse-to-fine multiresolution capacity to match and search, which Gabor transform don’t have. In addition, it has higher robust for light changing and could reconstruct 3D point cloud information densely. Finally, subdivision surface fitting technique is applied to get object-level local refinement of the surface reconstruction.A passive binocular stereo vision system with adaptive baseline is introduced. Here, "adaptive" means the length of baseline can be drived adaptively by SCM (Single Chip Micyoco) according to sample distance and required accuracy. Since adjusting baseline and shakable shooting may change the stereo calibrated parameters, a new semi-automatic calibration technology is put forward, which can gain new real-time calibration accurately. Furthermore, in order to improve the synchronization sampled by traditional dual-camera, a new passive binocular stereo vision system based on single camera is presented, as well as corresponding calibration algorithm, which solves the auto-calibration problem of non-pinhole model.There is also some work on the convergence analysis of subdivision method. It is well-known that the p-norm joint spectral radius is defined by a bounded collection of square matrices with complex entries and of the same size, which is used as convergence criterion of subdivision rules. The p-norm joint spectral radius for integers is investigated here, as well as some basic formulas and a simple proof of Berger-Wang’s relation concerning the∞-norm joint spectral radius. In addition, estimate of the eigenvalues of subdivision stochastic matrices is studied here, as well as some search algorithms from graph theory.At last, an improved 3D face recognition method as well as a new binocular stereo vision system based on single camera are proposed in this paper. Under the assumption that face is symmetrical, the point cloud is optimized automatically by filling holes and correcting. Then, simplified CANDIDE-3 model is used as initial subdivision controlling mesh, refined locally and levelly fitted. Meanwhile, Geodesic mapping technique is applied to normalize different expressions and face database is built respectively. Furthermore, pyramid structure is employed to compare and recognize 3D faces, which is also suitable for reverse seeking 3D face information. Experiments show that the new stereo vision system not only improves reconstruction accuracy, but also avoids robust decreasing caused by non synchronous shooting of two cameras. Moreover, subdivision surfaces used as storage can save space and provide theoretical support for comparison. Considering its low cost, the system is feasible to spread in many fields.更多还原