【作者】 傅荟璇；

【导师】 刘胜；

【作者基本信息】 哈尔滨工程大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 图像自动目标跟踪定位技术是世界各国精确视觉系统亟需解决的重要难题,是当前科技领域新兴的研究分支,也是计算机视觉领域的研究热点。尤其在某些场合,例如人无法参与或者工作量太大的情况下,研究智能化的无需人参与或仅需少量交互的系统是必然趋势。本文深入研究了图像处理领域中高精度的角点检测、双目立体视觉摄像机标定、目标识别与跟踪算法、定位等立体视觉系统研制开发的各种关键技术,从理论和实践上深入地进行了讨论,提出并构建了一个完整的光电目标跟踪定位立体视觉系统方案。各章的具体内容安排如下：图像特征点的提取是进行摄像机标定的前提,其提取的好坏直接影响着摄像机的标定精度,是顺利进行立体视觉和三维重建的基础。本文采用基于Harris角点检测原理的亚像素角点检测方法,利用角点邻域内图像灰度梯度变化与角点到邻域内任一点的矢量点乘为零的性质,采用迭代算法,获得了精度优于0.01个像素的亚像素角点坐标,解决了摄像机标定中高精度控制点坐标的获取问题。针对摄像机标定过程中复杂的成像和畸变模型,提出了基于最小二乘支持向量机的摄像机标定方法。利用最小二乘支持向量机来直接学习图像信息与三维信息之间复杂的非线性映射关系,不需确定摄像机具体的内部参数和外部参数,只需要知道部分已知点(已知世界坐标的点)的图像坐标。在双目视觉的情况下,两摄像机的位置关系不需具体求出,而是隐含在网络中。由于核函数的参数以及惩罚因子是影响最小二乘支持向量机性能的关键因素,采用AGA算法自动选择确定LS-SVM参数,建立了基于AGA-LS-SVM的摄像机标定模型。仿真结果证明这一方法对于提高标定精度的有效性。采用Haar特征作为目标识别的特征参数,解决目标自动识别与跟踪中存在的实时性、准确性的难题,将AdaBoost学习算法训练分类器,应用基于Haar特征的目标识别进行目标识别,试验结果证明该方法具有较好的实时性和较高的正确检测率。将Kalman滤波预测方法融入到Mean Shift跟踪算法中,利用Kalman滤波预测目标位置,有效地解决了背景中大面积颜色干扰的问题,提高了跟踪方法的抗干扰能力和速度。并以仿射变换来描述目标尺寸的变化,利用连续两帧中匹配窗口的最大相关系数确定最优匹配窗口搜索目标。当目标被遮挡时,利用目标先前运动轨迹采用LS-SVM对目标运动趋势做预测,在目标被遮挡后重新出现时能够重新找到目标,进行跟踪。跟踪试验表明,算法在遮挡、背景与目标颜色相近和目标尺寸变化等复杂情况下都能对目标进行准确跟踪,具有良好的实时性和较强的抗干扰能力。建立了目标自动识别与跟踪定位系统。进行了一系列目标自动识别与跟踪定位试验,试验结果表明：本文建立的目标自动识别与跟踪定位系统具有良好的实时性和较好的定位能力。更多还原

【Abstract】 Automatic object recognition and tracking in image is the key technology of the preciselyguided weapons, which need to be solved urgently in the world, is the current tech area emerging research branch, it is also an active field in computer vision. Especially in some situation, for example the human is unable to participate in or in the work load too big situation, the research intellectualization does not need the human to participation or needs the few interactive systems is the inevitable trend.In this paper deep research image processing domain high accuracy corner detection, camera calibration for stereo vision, object recognition tracking algorithm, location and so on stereoscopic vision system development each kind of key technologies, has carried on the discussion thoroughly from the theory and the practice. Proposed and has constructed a complete photoelectric object tracking location stereoscopic vision system plan. Each chapter of actual content arrangement is as follows:Image characteristic point extraction is required to achieve camera calibration, extraction quality immediate influence following calibration precision, and carries on the stereo vision and the three dimensional reconstruction foundation smoothly. This paper uses sub-pixel corner detection method based on Harris corner detection principle, using iterative method and the corner property that any vector from the true corner to a pixel point in the comer neighborhood is always orthogonal to the gradient vector of the image at the point we obtained corners subpixel coordinates whose precision precedes 0.01 pixel. This solves the problem how to achieve the control points coordinates with subpixel accuracy on camera calibration.In view of camera calibration process in complex imaging and distortion model, proposed based on LS-SVM camera calibration method. Least Squares Support Vector Machines are used to learn the relationships between the image information and the 3D information. It neither requires an accurate mathematical model nor needs any prior knowledge about the parameters. Only needs to know the image coordinate of partial known points(known world coordinate points). In binocular vision, the relationship between the location of the two cameras obtained without specific, but implicit in the map relations. The nuclear function parameter and penalty parameter is a pivotal factor which decides performance of LS-SVM, Uses the AGA algorithm automatic selection to determine the LS-SVM parameter, Has established the camera calibration model based on AGA-LS-SVM. Simulation results showed the validity to improving the calibration ccuracy.Uses the Haar characteristic to take the object identification the characteristic parameter, in the solution automatic object recognition and the track exists real time, accurate difficult problem. Using AdaBoost learning algorithm training classifier, apply the Haar characteristic carries on object identification, the experiment result show that the method has the good real-time performance and the high correct detection rate.Integrates the Kalman filter forecast technique to the Mean Shift tracking algorithm, using Kalman filter prediction objective position, has solved in effectively the background the big area color disturbance question, improves the tracking method anti-interference ability and speed. And describes the object size change by the affine transformation, matches the window using continuous two frame in the maximum correlation coefficient determination the optimal matching window search object. When the object is occlusion, using the object previous track forecast moving tendency with the LS-SVM, when the object is covered reappears could find object anew, carries on the tracking. The tracking experiment indicated that the algorithm in the occlusion, the background and the object color are close and the size change and so on the complex situation can carry on the accurate tracking to the object, has the good real time performance and anti-interference ability.Established the automatic object recognition tracking and location system. Has carried on a series of automatic object recognition tracking and location experiment. The experiment results shows that this papaer established the automatic object recognition tracking and location system has the good real time nature and location ability.更多还原