基于点云的复杂曲面物体3D建模关键技术研究

Aimed at the inadequate existing in traditional 3D modeling technology of the curved surface objects, this item will systematically research the following key technologies: point cloud registration，point cloud segmentation, curved surface fitting and features extraction:.(1) ZNCC(Zero-Mean Normalized Cross-Correlation Coefficient) is drawn into to measure the similarity degree of adjacent region curvature of the points pair, then the corresponding points are searched according by the adjacent region mark. And ICS（Iterative Corresponding Surface）algorithm is put forward which using local curved surface patches replace discrete points moreover using ADF(Adaptive Distance Function)replace geometry distance from a point to its corresponding curved surface patch, to build no-linear least square optimizing model and its resolving strategy;.(2)According to IFT(Implicit Function Theory),it is put forward that the automation segmentation method of point which based on ACM(Active Contour Model),and it is build that the energy function optimizing model based on the mean curvature of point then boundary extraction and region divide of point is realized by the self-adaptive topological evolution] of spatial curve. This method can avoid reconstruction of triangle grid and improve segmentation efficiency maximally under the foundation of guarantee segmentation precision, and it can remove the effect of point shortcoming effectively then realize high efficiency segmentation of point;.（3）It is researched that the method to fit general quadratic surface by disorder coordinate point then evaluate curvature tensor, and it is draw into that the curve surface fitting method of TLS3L（Total Least Squares on the Three Level Sets）in order to realize rapid and reliable fitting according to disorder point. Finally it is explored that the extraction method of invariant feature of the curve patches in the circumstance of Euclidean transformation].These research mentioned above can enrich the modeling theories and technologies of Laser scanning data.

针对传统曲面物体3D建模存在的不足，系统研究点云配准、分割、曲面拟合与特征提取等关键技术：（1）引入归一化互相关系数ZNCC衡量点对邻域曲率相似度，由邻域标识寻找对应点。提出 ICS算法，用局部曲面片代替离散点，并用自适应距离函数ADF代替点到对应曲面片的几何距离，建立拼合的非线性最小二乘优化模型和求解策略；（2）根据隐式函数理论IFT，提出基于活动轮廓模型ACM的点云自动分割方法，建立基于点云平均曲率的能量函数优化模型，通过空间曲线的自适应拓扑演化来实现点云的边界提取和区域划分。避免三角网格重构，在保证分割精度的基础上极大提高分割效率，有效去除点云缺陷的影响，实现点云的高效分割；（3）研究一种从散乱坐标点拟合一般二次曲面，然后估计曲率张量的方法，引入TLS3L曲面拟合方法，以实现对散乱点的快速可靠拟合；探讨欧氏变换下曲面片不变量特征的提取方法。本研究丰富了激光扫描数据的建模理论和技术。

基于激光点云的复杂曲面物体3D建模是计算机视觉和机器视觉领域研究的热点、重点和难点问题之一。本项目主要研究内容如下：.1、归一化互相关系数与迭代最近曲面片点云配准.针对无附加信息的激光点云数据，基于匹配点对衡量准则与迭代最近曲面片( ICS) 算法提出一种新的配准方法。引入归一化零均值互相关系数衡量点的邻域曲率相似度，构造一一对应的初始匹配点对有效数组，利用四元素和线性最小二乘法计算初始配准参数。通过局部曲面片代替离散点，建立参与ICS 算法的有效点集，并用一次近似距离代替点到对应曲面片的几何距离，建立配准的非线性最小二乘优化模型和求解策略。实例结果表明，与迭代拼接算法相比，该方法具有多视角普适性，且高效精确。.2、适应性距离函数与迭代最近曲面片精细配准.针对复杂曲面物体多视角激光扫描点云数据，提出了一种从深度图像到完整几何模型的配准方法。根据空间点相对位置在刚体变换下的不变特性，用曲率不变特征和归一化零均值互相关系数构造有效的初始匹配点对数组，基于单位四元数对匹配的特征点对进行坐标变换求解，完成数据粗略配准；探讨改正系数 的确定方法与步骤，计算不同改正系数下的均值误差，得到最佳改正系数 ；运用适应性距离函数和改进迭代最近曲面片精细匹配技术,将不同视角点云在三维空间进行最优化匹配；根据匹配结果计算配准误差，并对配准精度和速度进行了统计分析。数值试验结果表明，该方法在保证配准精度的前提下能有效提高配准效率。.3、参数活动轮廓模型距离图像分割.分析传统距离图像分割方法的缺陷，提出了新的基于参数活动轮廓模型的距离图像分割方法。首先阐述距离图像区域分割与活动轮廓曲线演化基本原理，以参数形式表达活动轮廓模型，并图示表达轮廓线上数据点的运动和点运动算法的伪码描述；其次，探讨内部能量、外部能量、图像能量和约束能量泛函组成，构造参数活动轮廓模型的完整能量函数，根据变分原理推导欧拉方程并进行离散化；最后在VC++6.0和MATLAB7.0相结合的开发平台下，利用乔里斯基因子分解求解。数值实验结果表明，与基于形态学水线区域的深度图像分割相比，该方法精确高效，效果良好，分割结果光顺连续，与人的主观视觉感知一致。.研究表明，复杂曲面物体激光点云配准，配准精度和配准效率不可兼得，同时提高计算效率和配准精度、深度图像分割定量评价指标体系构建、分割效果与分割效率双赢是下一步努力的方向。