【作者】 蔡勇；

【导师】 肖建；

【作者基本信息】 西南交通大学 ， 电气系统控制与信息技术， 2006， 博士

【摘要】 在反求工程中，基于点的点云处理技术是随着数据测量技术的进步而迅速发展起来的一门新兴技术。该项技术以点作为曲面绘制和造型的基本元素，在提高模型绘制与重建的速度、加强处理超大规模点云的能力和简化计算量等方面体现出独特的优势，目前已成为反求工程的一个研究热点。论文主要研究点云优化子采样曲面表示、图像修补、点云建模、核机器学习方法等内容，提出了几种基于核机器的点云处理方法，为该领域的研究工作提供了参考。 在已有研究成果的基础上进一步完善了基于点云的优化子采样曲面表示方法。对输入稠密点云采样形成稀疏圆或椭圆油彩，采样数据具有空间信息。用贪婪法选择曲面，用整体优化法消除漏洞，持续逼近输入数据重建曲面。在一定公差下可用较少的曲面实现无漏洞曲面逼近，产生的表面流畅、无漏洞、采样密度适中，并且可有效控制误差在规定的公差范围之内。理论分析与实验结果表明该方法具有良好的性能。 与统计学习理论相结合，提出了一种基于支持向量机的优化子采样曲面表示方法。讨论了基于点云数据的曲面表示问题，采用ε-支持向量回归机和ν-支持向量回归机实现点云数据的轮廓构建与曲面拟合，最后使用贪婪算法求解得到优化子采样，即一个曲面表示，它具有光顺性好，速度快等优点，展示出所提方法的优越性。 对图像修补的特点进行了仔细分析，在此基础上，提出了一种基于核机器方法的图像修补算法。对待修补区域的每一个像素点，首先取其八邻域色彩数据，形成一组不完整的曲线，然后用核机器对它们进行回归处理完成修补。实验对比了几种方法和几种核函数的性能，结果表明，论文中所提方法可以达到令人满意的效果，具有实用性。 分析了景物图像修补的特点，在此基础上构建了一类景物图像的样本数据集。将基于模式识别的核分类机引入该领域，首先采用支持向量机对待修补图像进行分类，然后利用匹配的模板对空洞进行边缘修补和色彩填充，最后根据空洞外部图像的细微特征对纹理和色彩进行二次修补。实验结果表明该方法的精度优于传统方法。更多还原

【Abstract】 In reverse engineering, advances in 3D scanning technologies have promoted the emergence and rapid development of point-based techniques. Points are primitive units of surface modeling and rendering, point-based techniques have become an important research field of reverse engineering. The particular advantages of point-based techniques are the efficiency at reconstructing and rendering very complex objects and environments, capability of dealing with dense scattered point cloud, and simplicity of rendering algorithms. In this paper, some important problems about point cloud, such as optimized sub-sampling curve plotting, hole filling, point cloud modeling, and kernel machine based on statistic learning theory and wavelet, have been analyzed and discussed. Several algorithms based on kernel technology have been proposed. Their feasibility has been shown by numerical experiments. The main contributions of this paper are following:An algorithm of optimized sub-sampling curve plotting has been proposed. Sparse circular or elliptic surface cells with space information can be extracted from dense point cloud by sub-sample. Optimized surface is found by greedy algorithm, and holes are filled by full optimized technology. Theoretical analysis and contrastive experiment results show following advantages of the algorithm: realization of full surface plotting without holes by fewer curves under given tolerance, good performance of smooth surface, well situated sampling density and robust controllability of error.Found on statistic learning theory (SLT), a sub-sampling curve presentation based on support vector machine (SVM) has been put forward. After discussions of surface figure, the lineament of point cloud and its imperceptible features have condtructed by ε-support vector machine and v - support vector machine. Greedy algorithm has been utilized to solve the geometry problems. Experiment results show that a smooth surface with good lubricity can be generated quickly.Based on the analyzing the features of point cloud hole filling technology, a kind of algorithm of hole filling based on kernel machine learning method is put forward. For each pixel of raw figure, color features of eight neighbors are distilled out to generate a group of incomplete curves. The holes on the curves are filled by regression based on kernel learning machine. Contrastive experiments of several algorithms and different kennels have been designed. The experiment results show that the algorithms proposed are satisfactory and could find their importance in更多还原