【作者】 张辉；

【导师】 张丽艳；

【作者基本信息】 南京航空航天大学 ， 航空宇航制造工程， 2009， 博士

【摘要】 物体的三维外形轮廓测量技术是实施逆向工程、产品质量检测、虚拟现实等的前端基础,在飞机、汽车、船舶、模具、娱乐、生物医学等行业有广泛的应用需求。本文深入研究了基于随机光场照射的双目立体测量关键技术,包括图像处理技术、摄像机标定技术、匹配重建技术、测量数据拼合技术和系统软硬件实现技术等,自主研发了RIMS(Random Illumination Measurement System)三维点云测量原型系统。本文的主要研究内容和成果可以概括为:1、提出了通过瞬时随机光场照射结合双CCD立体图像进行快速三维轮廓测量的总体方案。不同于大多数在一个时间序列上多次投射不同的结构化光照模式到被测物体表面的方法,本文提出的方案只需要一个自然光照下的立体图像对用于多视角测量数据拼合,以及一个瞬时随机光场照射下的立体图像对用于产生三维点云数据。其优点是随机光照立体图像对在瞬间拍摄,因此不会由于单次测量过程中被测物体的不稳定影响测量,非常适合现场测量和对非静态物体的测量。同时,该方案对光场投射装置没有严格的畸变度限制,而且不需要光栅移动(或转动)的机械装置,也无需散热结构,从而可以使投影器乃至整个三维测量系统紧凑小巧、成本降低。2、研究了图像的轮廓提取和椭圆中心提取算法,针对研究中采用的圆形目标点,改进了一个基于亚像素边缘的灰度矩最小二乘椭圆中心定位算法,与传统的灰度矩算法相比,有效地提高了圆形目标中心的图像定位精度。同时,针对在实际测量图像中可能存在的各种类型的伪椭圆目标,提出了伪目标的多准则过滤去除算法,能够高效稳健地识别出真实的特征椭圆,为系统标定及多视角测量数据拼合奠定了基础。3、深入研究了摄像机模型和双目立体测量系统的标定技术,通过实验,确定了RIMS测量系统的摄像机畸变模型,提出了一种简便易行的双目立体测量系统三步优化标定方法。该方法将基于平面模板的标定方法和自标定方法相结合,在传统两步优化法的基础上,进一步将标定板上特征圆点中心的空间坐标作为优化变量进行优化求解,并用已知两个点之间的实际距离精确恢复由此带来的系统绝对尺度变化。由于本文提出的三步优化标定方法考虑到了标定板几何误差的影响,并通过算法优化确定实际的标定板几何信息,从而降低了标定板的制作和计量校准要求,并且能够获得较高精度的标定结果。4、提出并实现了面向表面点云测量的双目立体图像匹配和三维重建算法。该算法在综合灰度信息约束和几何信息约束的基础上,利用最小二乘法,在同名像点匹配的同时得到相应三维坐标。提出了采用加权匹配窗口的方法,能够一定程度上改善模型细节部分的重建效果;采用独立于最小二乘模型之外的灰度矫正策略,一方面提高了算法收敛的稳定性,另一方面有助于提高算法效率;利用几何一致性和连续性约束,提出了用于确定匹配初始值的生长法,能够明显减少算法迭代次数,提高算法效率。实验证明,在辅以随机光场照射的情况下,本文方法能够根据一个立体图像对获得高质量的物体表面点云数据。5、提出并实现了基于圆形标记点的多视角测量数据拼合算法。首先研究了圆形标记特征在立体图像对之间的迭代松弛匹配算法,提出了“自适应部分赢者通吃”的更新策略;在多视角测量数据的拼合中,首先利用欧氏变换对长度的不变性,进行两两视图拼合,然后再对多视角测量数据的整体拼合结果进行全局优化,有效地减小了全局拼合误差。6、在基础理论和关键算法研究的基础上,构建了基于随机光场照射的双目立体测量原型系统RIMS的硬件平台,提出了软件的总体架构、数据结构以及数据管理体系,用VC++实现了文中提出的所有算法,实现了标定、测量、拼合等系统功能。文中提出的各项算法和系统功能均进行了翔实的实验、对比和分析,应用研发的原型系统,测量了大量具有代表性的三维实物,对测量精度和测量效率等方面进行了验证,效果良好。更多还原

【Abstract】 3D profile measuring is a highly demanding technique for industrial applications, such as reverse engineering, product quality control and virtual reality. It can be widely used in aerospace manufacturing, automobile making, shipbuilding, machine-tool industries, biomedicine, die manufacturing, etc. Among the existing 3D measurement techniques, image-based methods have demonstrated its vitality and promising future for its characteristics of non-tactility, high precision, efficiency and practicability. One optical method based on instantaneous random illumination has been fully studied in this thesis, and so does the key techniques for binocular stereo measurement including image analysis, camera calibration, matching, reconstruction, registration, software & hardware implementation, etc. The thesis’s main contributions are as follows.Binocular stereo measurement key techniques based on instantaneous random illumination have been researched. To acquire 3D point cloud data of a target object with low cost and simple setup, an instantaneous random illumination projection facility has been made to instantaneously project a random image pattern onto the object surface. Stereo images taken from two calibrated cameras with fixed geometry are matched and reconstructed for 3D point cloud acquisition. Since the stereo image pair is snapped simultaneously and instantaneously, the system is practicable for field measurement and moving objects’measurement because of its immunity to vibration of the object. In addition, since there are no strict constraints on light pattern projection such as distortion control, no raster movement mechanics and no cooling hardware, the structure of the projector and futher even the whole system can be compact and of low cost.Since the inputs of the binocular measurement are all images, image processing has been researched including contour detection and ellipse location. The outputs of the image processing algorithms serve as the input of afterwards camera calibration and image matching. Circular features are used as fiducial markers. Subpixel ellipse center detection and location algorithms based on least squares methods are analyzed. Since many kinds of fake features may exist in the real images, a fake ellipse removing algorithm is proposed. This ellipse extraction algorithm is robust and of high precision, and serves well as the input for the following stereo calibration and stereo registration algorithms.Camera model and binocular calibration have been researched. The measurement system’s camera distortion model has been carefully selected based on many experiments’analyse, and a practicable 3-step stereo calibration algorithm is proposed. In this method, planar plate based calibration is adopted. A normal 2-step calibration process is adopted which is similar to other typical calibration methods, and a further optimization procedure is included which takes feature points’3D coordinates as parameters. And after that, real metrics are obtained by scaling with respect to known distance between two fixed points. Since 3D coordinates’errors are taken into consider, calibration could be done with good precision while calibration plates’can be easily made.Aiming at binocular stereo reconstruction, a stereo matching and reconstruction algorithm is proposed. During reconstruction and matching, image intensity and geometrical constraint information are taken into account. The least squares method is applied to iteratively resolve the affine parameters and 3D coordinates. Image grey level correction is done before the least squares resolvement for the sake of speed. And with the assumed continuity constraint, a region growing strategy has been adopted. It is demonstrated by experiments that our matching and reconstruction algorithm can acquire fairly good surface point cloud for normal usage.Aiming at the registration of point cloud data computed from different views, registration algorithm with global optimization based on circular features has been proposed. Since the coordinate system changes between different views, a registration algorithm for transforming these point cloud data is needed. Firstly, an adaptive relaxation algorithm with“adaptive some winners take all”strategy is proposed for feature based image registration; secondly, 3D registration for two views is performed based on the fact that the Euclidean distance between any two points remains unchanged in different views; and at last, global optimization is implemented for the final 3D registration. Experiments show that the registration algorithm can robustly transfrom different views’data to a universal coordinate system.A prototype binocular measurement system, namely RIMS (Random Illumination Measurement System) based on instantaneous random illumination is designed and made with above key algorithms and techniques. For the hardware, suitable cameras and lens are selected; an instantaneous random illumination projector is made for instantaneous random pattern projection; a circuit board is designed for synchronization of camera image capture and lamp flash. For the software, stereo calibration, image matching and 3D reconstruction, cloud polygonization and display algorithm et al. have all been coded, implemented and integrated with VC++ for the primary functionality of the system.The proposed algorithms and system functionalities have all been fully experimented and analyzed. It has been demonstrated that the representive target objects’surface point cloud can be obtained fairly well by the prototype system. Because of its characteristics of effectiveness and good precision, the solution proposed in this paper has great potential tobe widely used in diverse applications.更多还原