【作者】 钟凯；

【导师】 史玉升； 李中伟；

【作者基本信息】 华中科技大学 ， 材料加工工程， 2013， 博士

【摘要】 三维面形测量技术目前已在航空航天、武器装备、汽车等领域得到了广泛应用,这些领域的核心零部件(如发动机缸体缸盖、叶轮叶片、进排气管道等)大多结构复杂、曲面不规则,在其设计制造过程中,采用三维面形测量技术不仅能够充分利用现有的设计制造成果,提高设计速度、缩短开发周期；还可精确控制产品质量,降低生产成本。然而,现有的三维面形测量技术大多只能测量静态物体,对于诸如叶片旋转过程中的动态三维形变、飞行器风洞实验中关键部件的动态三维形变、机车冲撞实验中机车表面连续变化的三维形貌和电磁成形过程中板料的连续三维形变等动态过程尚缺少有效的实时三维面形测量方法。这些动态过程的连续三维面形测量数据可与计算机模拟结果进行对比,以验证模拟结果的正确性,并有助于描绘和分析物体运动过程中表面形态的变化规律,为研究变形机理和改进相关工艺参数提供基础数据,具有重要的理论和应用价值。本文针对基于多视相移框架的动态物体三维面形测量技术中的系统非线性误差校正、基于多视相移框架的对应点查找、高精度的多视系统标定和大规模细粒度并行计算等关键技术的研究基础上,通过系统硬件精确同步采集设计与软件设计,开发一套全时间分辨率、全空间分辨率不连续具有任意复杂形状的动态物体三维测量系统。在以三步相移技术为基础的动态物体三维面形测量系统中,相位计算精度容易受到系统非线性响应的影响,使测量结果产生较大的周期性误差。为了提高三步相移的计算精度,对系统的非线性响应进行了系统的建模。该模型充分考虑了投影仪空间离焦现象的影响,能准确地揭示光栅图像空间非正弦化的规律。在此基础之上,提出了一种系统非线性响应伽马参数标定算法,该算法将标定的伽马值预先调制到预投影的光栅图像中,间接的对相位误差进行补偿。实验结果表明,该方法能将最大相位误差控制在0.045rad,从而有效地抑制由系统非线性响应引起的周期性测量误差,并且为整个三维计算过程打下了良好的基础。实现全时间分辨率、全空间分辨率任意复杂形状物体三维面形测量在科学研究与工程应用中具有重要的意义。为此,提出一种基于多视相移框架的对应点查找方法,该方法根据多视约束与相移原理,使得每个像素在解相完成后可以独立进行图像匹配与三维重构,无需相位展开过程。因此可以使测量系统实现任意复杂形状物体的全空间分辨率的三维测量。在连续测量的图像数据流中,任意三幅图像都可以用来重构三维数据,所以三维测量速度可以和相机采集速度一致,从而实现全时间分辨率的动态三维面形测量。计算过程中,根据相位图快速、稳定的图像匹配是保证动态物体三维面形测量稳定性的基础。匹配过程中,多光栅数与相位误差很容易导致误匹配,为此提出一种立体匹配两步方法(包括视差范围约束与混合一致性检查)来去除误匹配点,从而保证了图像匹配的正确率。此外,为了使匹配精度对相位误差不敏感,采用边界点相位校正来辅助边界点的对应点查找,同时利用极线约束对正确的匹配点进行精度优化。系统标定的精度直接决定了三维面形测量的精度与稳定性。为此,提出一种高精度的多目视觉标定方法。该方法主要包括单相机的内部参数标定与多相机间的位姿关系标定。针对单相机标定,采用基于平面标靶的柔性标定方法,得到相机参数的初值,然后利用光束法平差策略对相机的标定参数与标靶特征的三维坐标进行同时优化,进而消除标靶误差对标定结果的影响。针对多相机间的位姿关系标定,将世界坐标系固定在某一个相机上,利用两两相机间的极线约束建立目标方程,进而优化得到相机间的位姿关系,最后对位姿参数的平移矩阵中的比例因子进行修正。此外,系统在使用过程中,相机间的位姿关系容易受环境或人为因素影响,从而发生改变。为了保证系统的测量精度。在现有位姿标定方法基础上,研究了一种增强的相机姿态快速标定方法。该方法只要求相机拍摄一幅标靶图像,便能准确的计算出相机的位姿参数。实验证明,以上标定方法在噪声环境下具有良好的标定精度与稳定性。大数据量的密集计算一直是实时动态三维测量的瓶颈。为了有效提高三维计算速度,采用基于大规模、细粒度并行计算模型,充分利用GPU并行计算的性能优势,设计出应用在常规主流芯片上的高分辨率、高精度、实时动态三维测量并行算法。计算测试表明,GPU中三维计算的运行效率比在CPU中提高400倍左右,并实现每毫秒32000个点的实时动态三维计算,从而为实时动态物体三维面形测量技术的工程应用提供较为重要的基础。在前述理论与技术研究的基础上,研制一套全时间分辨率、全空间分辨率具有任意复杂形状的动态物体三维面形测量系统设备。硬件方面,设计了投影仪与相机的同步时序。针对其中的DMD翻转延时与相机接受到触发信号至感光响应的延时,设计两组测试实验分别测定了延时的大小,实现投影仪与相机精确同步,实验表明该系统能实现每秒220帧的精确同步采集。软件方面,根据前述的关键理论与技术开发出动态物体三维面形测量软件系统,实现了每秒220帧三维面形测量,测量分辨率为1024×768,并且探索了该系统在材料变形、运动追踪、人脸测量和手势与姿态识别中的应用前景。更多还原

【Abstract】 3D shape measurement technology has been widely used in the aerospace, weapons, automobile and other fields. The shape of key parts of these areas (such as the engine block, blades, exhaust pipes, etc.) is usually very complex, and the3D shape measurement technology not only be able to take full advantage of the existing design to improve the design speed and shorten the development cycle, but also control the product quality and reduce production costs. However, most of the existing3D shape measurement techniques can only measure the static objects. Dynamic3D measurement is increasingly important in many applications such as measuring the deformation of the rotating blades, the aircraft wind tunnel experiment, collision experiments and electromagnetic forming process. The dynamic process of continuous3D shape measurement results can be used to verify the computer simulation results, help to delineate and analyze the variation of the surface morphology and improve manufacture process, so it is very important in theoretical and practical researches.This paper presents a novel sinusoidal grating projection-based dynamic3D measurement techniques based on the development of nonlinear error correction technique, multi-view phase-shift framework based dynamic3D measurement, high-precision multi-view system calibration and massive fine-grained parallel computing. By accurate synchronization acquisition hardware design and software design, a high accurate, full temporal and spatial resolution, dynamic3D measurement system is developed.Typically, the phase value accuracy calculated by the three-step phase-shift algorithm will be influenced by the non-linear response of the system. In order to improve phase calculation accuracy, the nonlinear response of the measurement system is modeled. Compared with the existing models, it takes the defocus of projector into consideration and can describe the non-linear distortion of the fringe images in space accurately. On this basis, a gamma parameter calibration algorithm is proposed, and the calibrated gamma value will be encoded into the fringe images before measurement task for phase error compensation. Experimental results showed that the phase error could be control within0.045rad, and the3D measurement error was effectively suppressed, and it laid a good foundation for3D calculation process.It is very important to achieve full temporal and spatial resolution3D measurement of arbitary shape objects in scientific research and engineering application. So a stereo matching method based on multi-view phase-shift framework is proposed for dynamic3D measurement. In this framework, benefited from the trifocal tensor constraint, each pixel can explore its corresponding point independently only in the wrapped phase-map and reconstruct a3D point in space, so arbitrary shape including the discontinuous surfaces or step-like surfaces can be measured with full spatial resolution. Moreover, as any adjacent three images per camera in the continuous capturing stream can be used to reconstruct one3D frame, the3D acquisition speed can be as fast as the camera capturing speed. So our method can also perform full temporal resolution3D measurements. To enhance the resistance to false correspondence caused by multiple fringe number and phase error, a two-stage strategy, using disparity range constraint and mixed consistency check successively, is proposed to reject the wrong candidates. Further, the accuracy of the corresponding points is refined by performing edge point adjustment and correspondence refinement, and makes the3D result insensitive to the phase error.System calibration accuracy is very critical to the measurement accuracy, and a multi-view system calibration method is proposed. The calibration procedure contains two steps:(1) calibrating the internal parameters of each camera individually (2) determining the pose of each camera. For single camera calibration problem, Zhang’s flexible calibration method is used to obtain the initial values, and these values are optimized by the bundle adjustment strategy to eliminate the influence of the inaccuracy of the target. Then the world coordinate is located on the camera coordinate, the pose of each camera is calibrated by epipolar constraint, and the calibrated translation vector has to be revised to remove the scaling factor. In addition, due to the environmental factors, the pose of the camera might slightly change, which will make the measurement result inaccurate. In order to grarantee the measurement accuracy, a fast and enhanced pose calibration method is used to determining the pose of each camera. Experimental results showed that this method is accurate and stable.The large amount of data intensive computing is always the bottleneck of real-time dynamic3D measurement. In order to effectively improve the speed of3D calculation, a large-scale, fine-grained parallel computing model, which take full advantage of the performance of GPU parallel computing, is used to design a high-resolution high-precision, real-time dynamic3D parallel computing algorithms in the conventional mainstream chip. The experiment showed that the computation efficiency in the GPU is nearly400times faster than that in the CPU. And it achieved real-time calculation of32000points per millisecond, which will provide an important foundation in real-time3D measurement applications.On the basis of the above theory and technology researches, a dynamic3D shape measurement equipment is developed. In this equipment, the synchronous timing of the projector with the camera is designed to correctly capture gray scale images. Two test experiments are designed to determine the time delay of DMD flip and time delay of camera receiving a trigger signal and the photosensitive response. The experiments show that the system can achieve precise synchronization acquisition of220per second. On the other hand, a dynamic3D measurement software system is designed to achieve2203D frames per second dynamic measurement with spatial resolution of1024x768. Finally application prospects of this system in material deformation, motion tracking, the face measuring and gestures and gesture recognition is also described.更多还原