【作者】 吴庆华；

【导师】 史铁林；

【作者基本信息】 华中科技大学 ， 机械制造及其自动化， 2013， 博士

【摘要】 随着产品质量要求越来越高，检测手段越来越丰富，对产品表面缺陷识别也越来越严格。当前许多产品除了需要获取表面缺陷的二维信息外，还要通过三维检测手段了解缺陷的深度或高度信息。三维视觉是获取物体表面点云数据的主要方法，借助于三维视觉检测，使得原来无法识别的和不明显的缺陷，具有更加可靠的识别效果,这对产品生产工艺的调整更具有指导意义。所以基于三维视觉的表面缺陷检测越来越受到企业关注。本文提出并研究了基于线结构光扫描的三维表面缺陷在线检测基本理论和方法。本文的主要工作和创新点如下：（1）建立了基于线结构光视觉传感器三维表面扫描的数学模型。首先建立了线结构光视觉传感器测量的一般数学模型，然后调整测量坐标系使其XOZ平面与结构光平面坐标系重合，并假定测量坐标系的Y轴与结构光视觉传感器的扫描方向一致，这样使得三维扫描系统的各坐标系间的关系得到简化；最后在考虑摄像机径向畸变的情况下，建立了满足实际应用的三维扫描系统的数学模型。（2）对已有的三维锯齿靶标标定方法中的特征提取算法进行了改进，并利用改进后的特征提取法对线结构光传感器进行了标定。针对立体靶标标定方法的不足，提出了一种基于三圆点平面靶标的线结构光视觉传感器的现场标定方法，建立了标定模型，设计了一种特征点映射方法，比较方便的实现了线结构光平面与摄像机坐标系的位姿关系标定。（3）针对在线检测实时性较高的要求，提出了一种全自动的基于方向场的加权重心结构光条纹中心线提取算法，并与改进后的Steger法进行了比较，该方法效率高，且对三维表面的颜色和材质具有较强的普适性。（4）提出了一种基于高度颜色映射的三维表面缺陷识别方法。首先将点云高度映射为彩色图像，通过二维的最近点迭代快速配准方法，实现被测产品与标准产品的高度彩色映射图像配准，再对两幅图像进行色差计算，进而识别出缺陷；和传统的点云配准与偏差计算方法相比较，本文提出的缺陷识别方法充分利用了已有的二维图像处理方法，运算速度快，能够满足实时在线三维表面缺陷检测的要求。（5）搭建了三维扫描缺陷检测实验平台，并在Visual C++2010软件开发平台下开发了检测系统软件。并在此实验平台上，进行了磁瓦崩边缺陷检测、标签粘贴气泡缺陷检测、模具钢丝位置识别的实验，实验结果表明本文检测方法具有较强的可行性，可以广泛应用于各种产品三维表面缺陷的在线检测。更多还原

【Abstract】 With the increasingly high demand for product quality, and the detection meansgetting rich, product surface defects detecting is becoming more and more stringent. Atcurrent moment, in addition to obtain the2D (two-dimensional) information of the surfacedefects, but also the depth or hight of the defects should be known by the3D(three-dimensional) detection method. The main method for extracting the3D surfacepoint-cloud data is3D vision measurement. By using3D detecting method, some defects,which are not obvious or cannot be recognized by2D detecting method, can be recognizedmore reliably. Moreover, more advising information for adjusting production produce canbe obtained through3D vision measurement. So3D surface defects detecting based on3Dvision measuring is getting more and more attention by enterprises. In this paper, thetheory and application of3D surface defect on-line detecting based on line-structured laserscanning are put forward and studied.The main research work and creative items are discussed in the following.Firstly, the mathematic model for the3D scanning system based on line-structuredlaser vision sensor is built up. A universal mathematic model for line-structured laservision sensor is built up first of all. But this model is difficult to be realized and should besimplified. So, the measurement coordinate system is moved and rotated to theline-structured plane coordinate system, and the Y axis of the measurement coordinationsystem is alone with the scanning direction forward. Thus, the relationship among themeasurement coordination system, the line-structured plane coordination system and thecamera coordination system is easy to be transformed. Finally, using line-structured laservision sensor in the case of the camera radial distortion, the simplified mathematic modelfor the3D scanning system is built up. This model is easy to be realized for on-linescanning surface defects.Secondly, a feature extracting algorithm for the existing calibration method based on3D sawtooth target is improved and used to directly calibrate the line-structured laservision sensor. Against the disadvantages of3D target calibration methods, a newcalibration method based on a2D plane target is proposed. On the2D plane target, thereare three sizes of circle points which are arranged according to a certain position order.The calibration model is built up, and the feature points mapping method between targetcoordination system and camera coordination system is designed. The position and pose relationship between the line-structured plane and camera coordination system can becalculated conveniently on field.Thridly, the automatic line-structured laser stripe center extracting method based onweighted gravity formula alone with direction field is put forward. Compared with animproved Steger extracting method the efficiency of this method is higher, more suitablefor all kinds of surface with different color and material.Fourthly, the3D surface defects online detecting method based on a mapping policybetween the Z level and color value is proposed. According to the mapping policy, the3Dpoint-cloud is projected to a color image, and then the color image is transformed androtated to registrating with a standard color image, which is mapped from a standardproduct, using a2D interative closest-point algorithm. Thus, the color difference betweenthe projected color image and the standard color image can be calculated. Through colordifference threshold, the defect can be recognized. Compared with a common directly3Dpoint-cloud registrating and difference calculating method, the efficiency of this method ishigher, and can be used to online detect the3D surface defects.Finally, the3D surface scanning and defects detecting hardware system is built up,and a detecting software system is programmed in Visual C++2010platform. By usingthis device, edge-breaking defects of the magnetic tile surfaces are recognized, labelpasting bubbles defects are detected, and position locating of the steel wire in a model isrealized. All experiments show that the3D surface defect online detecting method basedon line-structrued laser vision sensor is feasible and can be used in all kinds of fields.更多还原