【作者】 周东生；

【导师】 魏小鹏； 滕弘飞；

【作者基本信息】 大连理工大学 ， 机械设计及理论， 2010， 博士

【摘要】 计算机视觉的基本任务之一是从摄像机获取的二维图像信息出发来计算三维空间中物体的几何信息,由此来重建或识别物体,并进一步指导机器认知现实世界。在上述过程中,摄像机标定是一个无法回避的步骤。在传统的摄像机标定领域内,研究者们在相关方面开展了广泛的工作,取得了丰富的理论和实践应用成果,包括摄像机模型的建立,摄像机标定物的构建,标定物的识别定位,标定物上特征点的提取,标定算法模型的构建,摄像机姿态判断以及基于标定技术的各种具体应用等。本文在国家自然基金等科研项目的资助下,针对上述部分领域内存在的一些问题开展了相关研究,主要的研究内容包括以下几个方面：1)提出了一种针对棋盘格标定物上特征点的自动提取方法。首先采用基于Gaussian平滑算子的改进型Wellner动态阈值算法实现了标定图像的二值化,然后采用Canny算子提取棋盘格的边缘信息,最后采用多次Radon变换与子图像划分相结合的方法,基于逐步推进,迭代求精的思路,实现了棋盘格上黑白相交处角点的自动提取,从而提高了摄像机的标定效率,并有助于实现一种全自动化的摄像机标定方法。角点提取过程中,采用图像分割与子图像处理的方法,在一定程度上纠正了由于镜头畸变带来的角点提取偏差。该算法提取的角点精度可以达到亚像素级,在存在遮挡的情况下仍然有能力预测到遮挡处的棋盘格角点坐标。实验表明,将采用本文方法提取到的角点信息应用于摄像机标定过程中后,最终得到的相机标定结果与采用现有成熟的人工交互方式的方法得到的结果精度相同。2)提出了一种基于角点分类的多棋盘定位方法。详细分析了非极大邻域、高斯方差和窗口、角点阈值以及权重系数等初始参数对Harris算子提取角点的影响,给出了针对棋盘格角点提取时的参数选取准则。构建了一种自适应旋转尺度和规模尺度的用于稀释棋盘格伪角点密度的二值化模板。综合运用密度聚类算法(DBSCAN)和模糊C均值(FCM)聚类算法完成了棋盘格伪角点过滤与多棋盘定位,并采用Radon变换算法对各棋盘边缘位置进行了精确提取。实验表明该方法对于具有复杂大背景的多棋盘定位具有良好的效果,为利用单张图片进行摄像机标定提供了一种行之有效的途径。3)采用摄像机标定技术实现了一种基于单目视频的低成本头部三维运动姿态跟踪与提取方法。该方法能够有效的提取头部姿态的6自由度(Degree of Freedom, DOF)运动参数。这些参数包括：分别沿X轴、Y轴、Z轴方向的平移运动参数Tx、Ty、Tz和分别绕X轴、Y轴、Z轴的旋转运动参数γ、β、α。在数据的后期处理中,采用了基于扩展卡尔曼滤波(Extanded Kalman flter, EKF)的算法对数据曲线进行了平滑和去噪,同时解决了由于缺帧等原因造成的祯间姿态跳跃问题。在实验中将采用本文方法提取到的数据与同步采用运动捕捉系统提取到的数据进行了比对,验证了本文方法的有效性和所得结果的准确性。4)提出了一种基于三维眼部几何模型的人眼瞳孔中心点定位方法。本文中首先基于人眼的生理信息和二维眼部图像,构造了人眼的三维几何模型。在此基础上,将提取到的瞳孔边缘信息从二维空间转换到三维空间,然后采用边缘拟合的方法完成了瞳孔中心点的提取工作。具体过程中,基于人眼瞳孔边缘点属于同一空间平面这一属性,本文采用最小二乘法拟合出该空间平面,然后利用该平面将椭圆拟合问题转化为简单的圆形拟合问题,从而减少了拟合瞳孔时边缘信息的需求量,降低了边缘拟合时的复杂度。实验表明,当瞳孔的遮挡面积达到75%时,采用该方法仍然能够准确的提取出瞳孔中心点的位置坐标。更多还原

【Abstract】 One of the basic tasks of computer vision is recognition and reconstruction of the 3D world by using the 2D image information which was acquired form the camera or any other image ac-quisition devices, and further guiding the machine to cognise the real world. Camera calibration is prerequisite to complete the task above. In the area of traditional camera calibration, researchers carried out extensive work and achieved rich theoretical results and applications, which including building the camera model, design and selection of the camera calibration targets, estimation of the camera pose, extraction of the feature points, build the camera calibration algorithm and achieved variety of specif c applications. Face to the existing problems in the above areas, and under the sup-port of the National Nature Science Foundation of China, this dissertation carried out systematic and in-depth studies. More details are as follows:(1). Automatic extracting the feature points from the chessboard targets. We f rstly used the Wellner adaptive threshold with a Gaussian smoothing operator to f nish the image binarization, and then used the Canny operator to acquire the edge information. In the end, we comprehensively used the multi Radon Transform and sub-image partition methods to f nish the whole automatic corner extraction process based on the route of coarse to f ne. Our work helps to realize a totally automatic camera calibration process, and will improve the efficiency of the camera calibration. We also used an image segmentation and sub-image processing method to reduce the corner location errors, and the results show that we achieved sub-pixel location accuracy. In the f nal, the experiments show that the calibration results which used our corners are equal to the results achieved by the existed mature approaches.(2). A multi-chessboards localization method was mentioned, which was based on the corner classif cation technique. First, we thoroughly analysis the function of the initial parameters belong to the Harris corner detection operator, respectively, and then the criterion of the parameter selec-tion were def ned. We constructed an adaptive rotation scale and size scale binary template, which used for diluting the density of the false corners. We comprehensively used the DBSCAN and FCM clustering algorithm to f nally f nish the false corners f ltering. In the end, the multi-chessboards border was extracted accurately by using the Radon Transform algorithm, respectively. The exper-iments showed that our approach especially suitable for locating multi-chessboards which have a wide and complex background, and also showed that our approach provided an effective way to calibrate a camera just by using a single image.(3). A low-cost and monocular video based 3D head pose tracking and extraction method by using the camera calibration technique was presented. The method can effectively extract the 6 DOF motion parameters which include three translation parameters:Tx, Ty and Tz following the axis X, Y and Z, respectively, and three rotation parameters:y,β, a surrounding the axis X, Y and Z, respectively. In the end, the Extended Kalman Filter (EKF) algorithm was used to solved the problem of data jitter, and some missing-data also could be repaired. We contrasted our method with the motion capture system, the results showed that the method is effective and has the capability of extracting the 3D head pose accurately.(4). Present a new human pupil center location method. We build a three-dimensional geo-metric model of the human eye by using the physiological parameters and 2D eye images. And then, the information of pupil edge was transform form 2D space to 3D space. Edge f tting method was used to acquire the pupil center. In detail, based on the 3D edge points of the pupil, we f rstly calculated the space plane by using the least square algorithm, and then we transform the ellipse f tting into circle f tting which could obviously reduce the demand of the edge information and the complexity of the f tting algorithm. The experiments showed that our method could still accurately extract the center coordinate of the pupil, when the block area arrived at 75%.更多还原