【作者】 袁野；

【导师】 欧宗瑛；

【作者基本信息】 大连理工大学 ， 机械制造及其自动化， 2002， 博士

【摘要】 计算机视觉作为当今最为活跃而又富有挑战意义的研究领域，其研究内容和应用领域相当广泛。本文以足球机器人视觉识别系统的开发为主要的工程应用背景，对边缘检测和轮廓跟踪算法、传统标定技术、自标定技术、隐式标定技术等相关技术进行深入研究。 边缘检测和轮廓跟踪在计算机视觉中有着重要的地位，利用小波变换良好的时频域局部化特性和多尺度特性，针对计算机视觉中需要提取单像素的点线等边缘的需求，本文提出了一种基于小波变换的边缘检测和轮廓跟踪算法，通过真正的边缘点是模的局部极值点这一特点，应用模糊算法对模极大值点进一步筛选得到单像素级的边缘，并在边缘跟踪时，有效利用原图像的信息，通过在小邻域内寻找局部峰值对丢失弱边缘进行了补偿。大量实验表明应用本算法进行轮廓提取和跟踪，可以提取单像素的边缘和连续的轮廓曲线。 计算机立体视觉中，首先必须解决的都是三维物点与二维像点的对应关系问题。因此，摄像机标定是计算机视觉实现的前提和基本问题。摄像机标定主要分为传统的标定算法和自标定算法。由于摄像机镜头存在着镜头畸变，对于要求精确定位的应用，需要进行畸变校正。传统的标定算法中各种优化算法存在着计算量大的不足，本文将单个自适应神经元应用到摄像机传统标定中，提出了一种考虑径向畸变和切向畸变的摄像机标定算法，它与其他优化算法相比具有简单实用、计算量小的优点。实验结果表明，该方法可以达到较高的精度。 不需要标定参照物的自标定技术正成为目前摄像机标定研究的重点，其中基于主动视觉的自标定算法可以通过控制摄像机做已知的运动而使问题简化，正在成为当今研究的一个热点。已有的基于纯平移运动的自标定算法对摄像机运动约束过多，且没有标定外参数。本文提出了一种基于纯平移运动的线性摄像机自标定算法，通过控制摄像机作三次不共面的纯平移运动，可以线性标定摄像机的内外参数。在此基础上还提出了一种考虑二阶径向畸变的非线性自标定算法，通过控制摄像机作四次不共面的纯平移运动，可以标定摄像机的五个内参数和两个径向畸变系数。这两种算法的优点是对平移运动的约束不多，解精度较高且具有一定的鲁棒性。 在需要很高的标定精度的情况下，基于精确数学模型的标定技术因为需要考虑各种非线性畸变因素，所以导出的畸变模型方程十分复杂。因此在三维测量应用中不依赖于确定的数学模型的技术——基于神经网络隐式视觉模型的标定技术将更有效。利用神经网络可以充分逼近任意的非线性关系且无须精确建模的特点，针对传统的立体视觉方法过程繁琐，对安装精度要求高的不足，本文提出了一种基于BP神经网络隐式视觉模型的立体视觉方法，该算法实施起来比较简便；针对已有的像差修正算法计算过程复杂的不足，提出了一种基于BP神经网络的修正成像误差的算法；针对具有共面特征的点的三维重构的应用，提出了一种基于径向基函数网络的二维平面测量算法。实验表明上述算法的精度较高，且鲁棒性较强。 本文成功的开发了足球机器人视觉识别系统，首先应用基于LVQ神经网络的颜色识别算法进行指定颜色属性的物体的识别，接着提出小球和机器人小车的识别算法，然后 大连理工大学博士学位论文应用i）；l练收敛后的神经网络进行摄像机隐式标定。实验结果表明，该算法的识别速度和识别精度都达到系统的要求。更多还原

【Abstract】 Computer vision is an active and challenging research field, its research area and applications area spread widely. Based on developing of vision recognition system of Micro-Robot World Cup Soccer Tournament, this dissertation is devoted to investigating some related techniques in computer vision, which include edge detection and contour tracking, conventional calibration techniques, self-calibration techniques, implicit calibration based on artificial neural networks.Edge detection and contour tracking are very important in computer vision. Because the single pixel edges are needed in computer vision, an algorithm of edge detection and contour tracking is proposed using the good local character and multi-scale character of wavelet transform in the dissertation. The fuzzy algorithm is applied to pick the model maximum points, so that the single pixel edge can be obtained. Edge image was tracked through using the image information sufficiently. Through finding local peak in a little line adjacent which is perpendicular to the edge direction, the missing edge can be compensated. The experimental results show that the single pixel edge can be obtained, and continuous contours can be extracted.In computer stereo vision, the primary problem to solve is the relations between the 3D points and the 2D image points. So the camera calibration is the premise and the basic problem. There are conventional calibration and self-calibration in camera calibration techniques. Since the cameras lens used in computer vision sustain a lot of nonlinear distortion, recent research efforts has been concentrated on the distortion correction techniques. The optimization algorithms in conventional calibration field have the shortcoming that the computing quantity is huge. In order to improve it, a single adaptive neuron algorithm was developed in conventional camera calibration in this dissertation, and a calibration algorithm considering radial and tangent distortion based on single adaptive neuron is proposed. Compared with the ordinary optimization algorithm of calibration, this algorithm gains simplicity, less computing quantity, and also keeping high accuracy .Camera self-calibration is becoming the important field of calibration research. Camera self-calibration based on active vision makes the problem simplified taking advantage of controlling camera to do known movement. The existed self-calibration algorithms based on translation motion have the limit in constraining the translation too strictly and can not get the external parameters. A linear camera self-calibration approach is proposed in this dissertation. The intrinsic and external parameters can be calibrated linearly by controlling the camera to undergo 3 translations or more which are not co-planar. Based on this algorithm a self-calibration approach taking account of camera two-degree radial distortion is proposed. The five intrinsic parameters and two-degree radial distortion coefficients can be calibrated by controlling the camera to undergo more than 4 translation motion which are not co-planar.inCompared with other methods of self-calibration, these algorithm gains simplicity, strong robustness and high accuracy.The equations derived are more complicated if more precise model is employed for high accuracy. The technique, which doesn’t need to have an explicit model - calibration based on neural networks implicit vision model, is more effective. Since BP neural network can implement any nonlinear relationship from input to output and needn’t to model, and the classical stereo vision approach based on explicit model are very complicated, an algorithm of stereo vision based on BP neural networks implicit vision model is proposed. This algorithm gains simplicity and convenience. The correction of camera distortion is a main part of camera calibration. An algorithm of camera distortion correction based on BP neural networks is proposed in this dissertation. In the special applications that the 3D points are coplanar, a 2D plane calibration algorithm b更多还原