题名基于图像序列特征点三维测量技术的研究
作者孙亦南
学位类别博士
答辩日期2005-04-27
授予单位中国科学院沈阳自动化研究所
授予地点中国科学院沈阳自动化研究所
导师王越超 ; 刘伟军
关键词三维测量 图像序列 特征标志 图像处理 计算机视觉
其他题名Research on 3D Measurement of Coded Targets Based on Image Sequences
学位专业机械电子工程
中文摘要三维测量技术是反求工程中对物体的三维几何形面进行三维离散数字化处理的方法。目前已有多种三维测量技术并得到了广泛应用,但这些方法仍存在着各种各样的使用限制条件。近年来,随着相关领域技术的不断完善以及硬件成本的降低,由图像序列进行三维物体测量技术日益受到人们的重视。但由于涉及到多个学科的相关知识,这种方法仍然存在着不少问题。本文针对这些问题,提出了一种面向快速原型制造领域的基于图像序列的特征点三维测量方法,并利用所开发的一套三维测量原型系统对一些技术方法进行了实验验证。本文首先分析了当前反求工程领域中数据获取的理论和方法,指出目前现有方法的不足。通过对射影几何、针孔模型成像理论及三维空间层次关系的研究,指出由图像进行三维测量的可能性,建立了面向快速原型领域的基于图像序列进行三维测量的理论框架。提出了一种基于特征标志组合辅助进行相机标定及三维重建的方法。针对计算机视觉中固有的匹配难题,设计了一套特征标志组合。针对获得图像的特点,提出了面向特征标志的图像分割算法流程,并采用了基于Freeman链码的跟踪算法将图像中的每个特征标志分割出来。在实际应用中证明了该算法具有较高的分割效率。根据特征不变量的定义,分析了三维空间下各层空间中特征不变量的关系,指出了利用仿射不变量来近似表达射影空间下特征不变量的可能性。针对本文特征标志的特点,提出了四种不变量作为特征标志的特征值,建立了特征标志的标准特征数据库。在此基础上,提出了一种基于不变量进行特征识别的识别策略,得到特征标志在每幅图像中的二维信息。实验证明,采用该方法能够有效的对特征标志进行识别。 在得到特征标志的图像信息后,需要将二维信息重构为三维信息,这就是摄像机标定的工作。本文系统研究了目前主要的摄像机标定方法,确定了使用相机自标定方法进行三维重建的技术路线。基础矩阵包含了图像间几何约束的所有信息,它的精确估计在摄像机标定中起着重要作用。为此,本文提出了一种基础矩阵的加权线性算法,该算法利用点到极线距离的几何特征作为每个点的加权因子,使匹配精度高的点对基础矩阵的影响更大,与传统的基础矩阵计算方法相比,该方法提高了计算精度。根据基础矩阵确定射影重构的初始框架,为了消除噪声的影响,采用了最优三角形方法对最初两幅图像中的匹配点进行初始重构。对每一幅新加入的图像,采用最优估计算法计算新的投影矩阵,并采用迭代扩展Kalman滤波技术对已有的重构点进行优化。在处理完全部图像后,采用集束调整方法对结果进行了全局优化处理。在得到高精度的射影重建后,需要将其升级到欧式重建。为此,本文系统研究了当前相机分层标定的理论和方法,确定了由射影空间—仿射空间—度量空间—欧式空间进行分层标定的技术路线。利用模约束理论进行特征的仿射空间重建,采用多面体-线性同伦延拓法求解模约束方程组提高求解的鲁棒性。为了避免数值计算稳定性的影响,采用平行线约束方法辅助进行相机标定,最后利用线性方法并对结果进行集束调整从而完成特征点在欧式空间下的三维测量工作。在研究并实现所需技术的基础上,参照目前在基于图像建模领域具有代表性的各种软件系统,结合快速原型制造领域的特点,利用VC++开发了一套基于Windows的具有人机交互功能的特征点三维测量原型系统。其中,设计了一个具有简单特征的圆形标记作为与后继测量工作中坐标系融合的纽带,结合图像的特点,提出了圆形标记的定位方法。同时利用对极几何约束提出了圆形标记的半自动匹配算法。通过一系列的实际实验表明,该系统初步具有了三维测量的主要功能。
索取号TP391.4/S98/2005
英文摘要The thesis proposes a method of 3D measurement of coded targets based on image sequence and experimentally testifies some technical methods by the 3D measurement prototypal system we developed. Based on the research of projective geometry, pinhole imaging model and the stratified relations of 3D spaces, the thesis points out the possibility of 3D measurement only used images. The thesis proposes a way to camera calibration and 3D reconstruction utilized a set of coded targets. For the purpose of avoiding the correspondence problem which is the biggest problem in the field of computer vision, a set of coded targets is designed and aiming at the features of images, the algorithm process is proposed which is to get the 2D coordinates of each coded target in each image. The tracking algorithm based on Freeman code is proposed to track each region of coded target. Based on the definition of the feature invariant, the relationship of feature invariants in different layer of 3D space is analyzed and the possibility of replacing projective invariants by affine invariants is also indicated. A recognition strategy is proposed using four different invariants to recognize each coded target. So the 2D information of each target in each image is gotten. A systematic research is carried out on the main methods of camera calibration and the self-calibration method is decided to reconstruct the 3D object. A weighted algorithm to compute the fundamental matrix is proposed. Using the distance between matched point and corresponding epipolar line as the weighted factor, the influence of high-matched point to the fundamental matrix is bigger than low-matched one. The computational accuracy based on this algorithm is improved comparing with traditional algorithms. The initial framework of projective reconstruction is defined based on the fundamental matrix, and the initial reconstruction of matched points in the first two images is also defined using the method of triangulation. For each new added image, the golden standard algorithm is used to get the new projection matrix of new image and the existed 3D reconstructive points are optimized using Iterated Extended Kalman Filter method. After getting all the projection matrixes, the results are optimized using the method of bundle adjustment. The thesis researches the current methods of stratified calibration and decides to calibrate the camera from projective space to affine space, then to metric space and to Euclidean space at last. The method of modulus constraint is used to get the affine calibration and the polyhedral-linear continuation method to solve the system of equations composed by modulus constraints. For avoiding the influence of the stability of the numerical computation, the constraint of parallel lines is used to assist the calibration. After all, using the linear method and optimizing the results by bundle adjustment, the 3D measurement in Euclidean space is finished. An initial system in Windows is developed which has the ability of human-computer interaction. A simple circle target is designed as the connection with following work to fusing two difficult coordinates systems. The method of segmentation of the target is proposed based on the character of images. Using the epipolar geometry, the semi-automatic matching algorithm for the target is proposed. A series of experiments show that the system elementarily has the primary functions of 3D measurement.
语种中文
产权排序1
公开日期2012-08-29
分类号TP391.4
内容类型学位论文
源URL[http://ir.sia.ac.cn/handle/173321/9550]  
专题沈阳自动化研究所_工业信息学研究室_先进制造技术研究室
推荐引用方式
GB/T 7714
孙亦南. 基于图像序列特征点三维测量技术的研究[D]. 中国科学院沈阳自动化研究所. 中国科学院沈阳自动化研究所. 2005.
个性服务
查看访问统计
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。


©版权所有 ©2017 CSpace - Powered by CSpace