基于立体相位展开的结构光实时三维测量及重构技术研究

发布时间：2018-03-17 05:19

本文选题：三维测量　切入点：相位展开　出处：《南京理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,基于相位测量轮廓术的结构光三维测量方法凭借其高精度、高效率和低成本的优点,被广泛应用于产品质量检测、虚拟现实和机器视觉等领域。随着社会需求的不断发展,获取动态场景的三维信息变得越来越重要,而目前大多数的测量方法还只能实现对静态物体的测量。相位测量轮廓术至少需要3帧条纹图像恢复出待测场景的相位信息,但是这样的相位大多被反正切函数截断在[-π,π)之间。想要得到无歧义的相位通常需要额外的条纹图像来辅助展开包裹相位,这无疑进一步加重了运动对测量的敏感程度,不利于实现对动态场景的测量。为此,本文提出了一种基于立体相位展开算法的实时三维测量方法,并采用该方法搭建了一套三维测量系统。1.针对系统的非线性响应导致的周期性误差,建立了系统的灰度响应函数,并采用查找表的方式进行校正;然后在此基础上,借助三步相移算法提取待测场景的相位信息,以及基于张正友标定法对双目--投影仪系统进行了标定,获取系统结构参数。2.为了在不增加条纹图像幅数的情况下获取绝对相位,本文提出了立体相位展开算法:引入一个额外的相机,通过遍历所有的条纹周期,分别将左相机上的像素点映射到三维空间后,再反投影到右相机上,最后利用图像匹配来判断该周期是否为像素点对应的包裹相位级次,从而实现相位展开;考虑到算法的实时性和准确性,在实现图像匹配的过程中,分别使用了极线约束、轴向测量范围约束和光强平方差约束逐步排除错点,最后利用像素点的邻域窗口相关计算来确定最终的匹配点;为了提高相关计算中每个像素点的标识度,选择在相移图案中嵌入散斑信息。区别于传统的在整幅图像上随机嵌入散斑的方法,本文选择在相移图案的灰度冗余区域嵌入随机散斑,在提高随机散斑灰度峰值的同时,也避免了条纹图像对比度和信噪比的降低。3.为了提高系统的实时效果,针对本文在算法实现上的特点:从相位提取、立体相位展开到立体重构,每个像素点的计算在所有算法步骤中都是独立的。因此,本文引入了基于图像处理器(GPU)的大规模数据并行计算架构的对算法进行优化。最终,系统成功实现了有效测量点数目为94,163时26帧/秒的实时三维测量,测量精度(RMSE)约为 0.1991mm。
[Abstract]:In recent years, structured light 3D measurement based on phase measurement profilometry has been widely used in product quality detection because of its advantages of high precision, high efficiency and low cost. With the development of social demand, it becomes more and more important to obtain 3D information of dynamic scene. At present, most measurement methods can only measure static objects. Phase measurement profilometry requires at least three frames of fringe images to recover the phase information of the scene to be tested. But most of these phases are truncated by the inverse tangent function between [- 蟺, 蟺]. In order to obtain an unambiguous phase, an additional fringe image is usually required to assist in unwrapping the phase, which undoubtedly further increases the sensitivity of motion to the measurement. This paper presents a real-time 3D measurement method based on stereo phase unwrapping algorithm. Using this method, a set of 3D measurement system. 1. Aiming at the periodic error caused by nonlinear response of the system, the gray response function of the system is established, and the method of look-up table is used to correct it. The phase information of the scene to be tested is extracted by the three-step phase shift algorithm, and the binocular projector system is calibrated based on Zhang Zhengyou calibration method to obtain the system structure parameter .2. in order to obtain the absolute phase without increasing the fringe image amplitude, In this paper, a stereo phase unwrapping algorithm is proposed: an extra camera is introduced. By traversing all the fringe periods, the pixels on the left camera are mapped to 3D space and then projected back to the right camera. Finally, the phase unwrapping is realized by using image matching to determine whether the period is the order of wrapping phase corresponding to pixel points. Considering the real-time and accuracy of the algorithm, the polar line constraints are used in the process of image matching. The axial measurement range constraint and the intensity square difference constraint are used to eliminate the error point step by step. Finally, the final matching point is determined by using the neighborhood window correlation calculation of the pixel points, and in order to improve the marking degree of each pixel point in the correlation calculation, Choosing to embed speckle information in the phase shift pattern. Different from the traditional method of embedding speckle randomly on the whole image, this paper chooses to embed the random speckle in the redundant region of the phase shift pattern, which increases the peak value of the random speckle intensity at the same time. In order to improve the real-time effect of the system, aiming at the characteristics of the algorithm in this paper: from phase extraction, stereo phase expansion to stereo reconstruction, The computation of each pixel is independent in all the steps of the algorithm. Therefore, this paper introduces a large-scale data parallel computing architecture based on image processor GPU to optimize the algorithm. The real time 3D measurement with effective number of measuring points 94 and 163.26 frames / s has been successfully realized. The measurement precision is about 0.1991mm.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】