基于多视图的三维模型重建方法研究

发布时间：2018-04-02 15:22

本文选题：三维重建　切入点：特征点匹配　出处：《山东大学》2009年博士论文

【摘要】： 三维模型获取是计算机图形学和计算机视觉领域的一个基本研究问题。然而,利用建模软件(比如3D MAX和Maya等)手工进行三维模型构建是十分繁琐和代价昂贵的工作。因此,研究如何从现实世界直接和快速地获取三维模型,成为该领域的热点问题。目前,基于现实物体的三维结构获取作为一种数字存储和记录技术,在物体建模、场景建模、真实感绘制、机器人导航、目标识别和三维测量等科学和工程领域以及考古学、广告、娱乐等其他文化领域有广泛的应用需求。基于现实物体的三维模型获取方法主要分为主动方法和被动方法。其中,主动方法以使用三维扫描仪的方法为代表。被动方法则指基于二维图像的三维重建方法。基于图像的三维建模方法具备低成本,灵活和能够直接获取彩色纹理等特点,是三维激光扫描等主动方法的有益补充。基于图像的三维建模方法主要分为基于标定图像和基于未标定图像两种方法。其中基于标定图像的方法需要在重建场景中预先放入标定物,具有时间和空间的限制性。基于未定标图像的三维建模方法仅依赖图像间的特征匹配关系,克服了基于标定图像方法的限制,具备良好的应用前景。目前,基于未定标图像的重建方法往往针对窄基线图像序列,这使得重建完整模型需要过多的图像数目,提高了时间和空间复杂性。本文基于多视图未定标图像的局部特征以及多视图之间的约束关系,以构建复杂完整的三维模型为目标,对三维重建的整个流程进行了深入研究。主要研究工作和创新点总结如下: 1.提出了新的图像特征点描述子提出了一种新的描述图像局部特征的方法。该方法首先提取图像中尺度不变的局部特征点,其次对特征点周围一定尺寸的邻域内梯度数据进行归一化处理,得到特征点像斑,然后采用独立成份分析(ICA)技术提取特征点像斑的独立成份,作为特征点的特征描述向量。该种描述子提高了局部特征的独特性和匹配精度,可用来解决宽基线多视图图像的特征点匹配问题,使得重建完整三维模型需要的图像数量较少,有利于降低重建工作的时间和空间的复杂性,为三维结构恢复奠定良好基础。 2.基于二维信息的三维结构和相机运动参数估计算法设计并实现了从二维图像空间重构三维空间的点云和相机运动参数估计的算法流程。 (1)提出了基于全局优化的基础矩阵求解方法。给出了一种新的使用全局最优技术,对基础矩阵进行非线性估计的方法。首先,在满足秩为2的前提下,使用最少变量对基础矩阵进行参数化。其次,为基础矩阵建立非凸的全局最优估计模型,并利用线性矩阵不等式松弛法转化非凸问题,使其最终可通过标准线性矩阵不等式(LMI)工具求解。最后,使用RANSAC迭代框架,基于最优图像距离误差,对求解的基础矩阵进行优化,进一步提高了结果的鲁棒性。 (2)提出了仅依赖基础矩阵精度的射影空间多视图递推公式,并基于此进行场景射影重建和度量重建。将射影空间投影矩阵形式化为统一的形式,基于基础矩阵和增量法,估计对应不同视图的投影矩阵。采用双视图估计,三视图局部优化,串联估计所有视图运动参数的策略,有效减少估计过程的累积误差。所估计的射影空间投影矩阵和同时重构的射影空间点云作为自标定算法的输入,标定出相机的内参矩阵,从而将投影矩阵和点云从射影空间升级至度量空间。由于基础矩阵的估计具备鲁棒性,因此,基于我们的方法所计算的相机投影矩阵,稳定性高,误差较小,使重构的点云具有良好的精确性。 3.提出了三维点云的优化算法。 (1)提出了基于SBA框架和随机行走模型的非线性优化算法。在对三维点云进行优化时,二维匹配点是优化算法的输入,采样精确的二维匹配点对提高优化算法的性能非常重要。提出一种各向异性的随机行走模型,用来重新采样图像空间匹配点。以重采样的匹配点对,投影矩阵参数和初步估计的三维结构为优化初值,利用SBA框架进行局部和全局优化处理。最后在RANSAC框架中进行迭代优化和最优参数选取。 (2)提出基于图像轮廓的点云调整方法。根据采样视点图像空间的轮廓数据,逆向修整三维空间的点云数据。首先,根据轮廓信息计算需要调整的三维点集合M,其次,提出两种方法,包括步长调整法和直接计算法对集合M中的点沿其内法向进行启发式调整。 4.提出了基于马太效应概率模型的多视图纹理映射算法。提出了基于多视图图像,针对复杂三维模型的自动纹理映射算法。获取三维结构的序列图像,作为纹理图像,映射至三维模型表面,以增强模型的视觉效果。在迭代框架中,基于马太效应法则,抽象出模型三角网格所属最佳纹理图像的变换概率模型,对所有输入的多视图纹理图像进行自动重采样,并对网格纹理分布进行优化,使纹理效果最优的同时使纹理接缝尽量减少。另外,提出了算法进行纹理接缝融合和纹理表面空洞修补。
[Abstract]:The 3D reconstruction is a basic research topic in the field of computer graphics and computer vision. However, by using the software (such as 3D MAX and Maya) manual construction of three-dimensional model is very tedious and costly work. Therefore, how to study from the real world directly and quickly obtain the three-dimensional model, has become a hot issue in the field. At present, the three-dimensional structure of the real object to obtain as a digital storage and recording technology based on object modeling, scene modeling, realistic rendering, robot navigation, object recognition and 3D measurement and other fields of science and engineering, archaeology, advertising, entertainment and other cultural fields have broad applications.
The 3D model of real objects acquisition method is divided into active and passive method. The method is based on the method, the active method is represented by three-dimensional scanner. The passive method refers to the method of 3D reconstruction based on 2D images. Image based 3D modeling method which has low cost, flexible and can directly get the color texture, is good of 3D laser scanning active methods.
Image based 3D modeling method based on calibration is mainly divided into two kinds of methods based on image and image. The calibration method based on Uncalibrated Image reconstruction needs in the scene in advance into the calibration object, with limited time and space. The 3D modeling method of uncalibrated image depends only on the feature matching between images based on the relationship, overcome the calibration method based on image, have a good application prospect. At present, the reconstruction method based on uncalibrated images is often for narrow baseline images, which makes the reconstruction of complete model requires the number of image too much, improve the time and space complexity.
In this paper, based on the local characteristics of multi view uncalibrated images and the constraint relationship between multiple views, the whole process of 3D reconstruction is studied in order to build complex and complete 3D models. The main research works and innovations are summarized as follows.
1. a new feature point descriptor is proposed.
A new method is proposed to describe the local image features. The method firstly extracts local feature points in the image scale invariant, then normalized to the neighborhood feature points around the size of the gradient data obtained feature points like spot, and then using independent component analysis (ICA) technique to extract feature points as independent component the spot, as the feature vector description. The descriptor improves the unique local features and matching accuracy, can be used to solve the multi view wide baseline image feature point matching problem, making the reconstruction of complete 3D model images need less, is conducive to reducing the complexity of the reconstruction work of time and space, lay a good foundation for 3D structure recovery.
2. estimation algorithm of 3D structure and camera motion parameters based on two-dimensional information
The algorithm flow of the estimation of the motion parameters of a point cloud and a camera from a two-dimensional image space is designed and realized.
(1) put forward the basis matrix solution method based on global optimization. This paper presents a new global optimal technology, nonlinear method for the estimation of the fundamental matrix. Firstly, to meet the rank 2 under the premise of using the least variable parameters of fundamental matrix. Secondly, based on the global optimal non matrix the convex estimation model, and convert the non convex problem using linear matrix inequality relaxation method, the final by standard linear matrix inequality (LMI) tools to solve. Finally, using RANSAC iteration scheme, the optimal image distance error based on fundamental matrix to solve the optimization, to further improve the robustness of the result.
(2) propose a multi view projective space depends only on the accuracy of the fundamental matrix recursive formula, and based on this scene projective reconstruction and metric reconstruction. The matrix form of projective space projection into a unified form, fundamental matrix and incremental estimation method based on projection matrix corresponding to different views. Estimated by the dual view, three view of local optimization, estimation of motion parameters of all view series strategy, effectively reduce the cumulative error estimation process. The estimated projection matrix and projective space and projective space point cloud reconstruction as a self calibration algorithm for the input, standard reference matrix camera set, which will be the projection matrix and the point cloud from projective space to upgrade to measure space. Due to the estimation of the fundamental matrix robust, therefore, the camera projection matrix, calculated by our method based on high stability, the error is small, the reconstruction of point cloud with good precision It's true.
3. the optimization algorithm of three dimensional point cloud is proposed.
(1) we propose a nonlinear optimization algorithm of SBA framework and random walk model. Based on the three-dimensional point cloud is optimized, the two-dimensional matching point is optimization algorithm for the input, sampling accurate two-dimensional matching points is very important to improve the performance of the proposed algorithm. The random walk model is an anisotropic, re sampling image the space matching point. By matching point resampling of the three-dimensional structure of the projection matrix parameters and preliminary estimates for the optimization of the initial value, the local and global optimization using SBA framework. Finally, iterative optimization and optimal parameter selection in the RANSAC framework.
(2) proposed adjustment method based on the point cloud image contour. According to the contour data sampling view image space, point cloud data in reverse dressing in three-dimensional space. Firstly, according to the contour information to calculate the 3D point need to adjust the set M, secondly, put forward two kinds of methods, including the step adjustment method and direct calculation method of the M collection the point along the inner method to heuristic adjustment.
4. a multi view texture mapping algorithm based on the Matthew effect probability model is proposed.
Propose a multi view image based on texture mapping algorithm for automatic complex 3D model. For image sequences 3D structure, as texture image is mapped to the 3D model surface model, to enhance the visual effect. In the iterative framework, based on the Matthew effect, get the probability model of triangular mesh model transform is the best texture multi view images, the texture image of all input for automatic resampling, and the grid texture distribution is optimized, the optimal texture and texture seams to reduce. In addition, this paper puts forward the method of fusion seam texture and texture surface patch the hole.

【学位授予单位】：山东大学
【学位级别】：博士
【学位授予年份】：2009
【分类号】：TP391.41

【引证文献】