多目立体视觉中图像匹配技术研究

【作者】 王静；

【导师】 张维忠；

【作者基本信息】 青岛大学 ， 计算机应用技术， 2010， 硕士

【摘要】 本文研究的主要内容是多目立体视觉的图像特征点匹配问题。图像特征点匹配指的是找出物体从不同角度拍摄后在不同图像上投影点的对应。为了提高图像匹配算法的精度和可靠性,首先将一套明显区别于背景物体的标记点粘贴到物体上对其进行拍摄,然后对拍摄图像进行处理,实现不同图像间同名点的匹配。为了使得特征点匹配以及三维重建的结果更加精确,本文对边缘检测和特征点匹配问题进行了研究。主要的研究内容及创新点如下：1)、提出了一种改进的OFMM算法。由于OFMM算法运算量大,对OFMM算法的模板进行了改进,使边缘判定条件发生了改变。2)、提出了一种新的标记点自动检测算法。为了使边缘检测算法达到高效、快速、准确的运行效果,首先使用Canny算法对图像进行初始边缘定位,然后根据标记点的特征对无效边缘进行过滤,其次使用改进后的OFMM算法对图像边缘进行进一步修正,再次对标记点使用改进的算法分类,最后确定标记点的亚像素中心坐标。3)、提出一种基于相机姿态的匹配算法。首先根据已经匹配的编码元获取各幅图像的相机姿态,然后在相机姿态已知的情况下,根据极线约束实现非编码元的匹配,最后为了增加匹配点对并提高匹配的准确性,使用对称性距离约束对初匹配结果进行精匹配。更多还原

【Abstract】 Image matching in multi-stereo vision is studied in this thesis. To improve the precision and reliability of matching of corresponding points between different images taken from different orientation by one single camera, a set of refrence points is applied as an assistant utility to measure the object. The main contents and contributions of this thesis are as follows:(1) An improved algorithm is put forward based on the OFMM. Owning to the great computational complexity of the OFMM algorithm, Canny operator is used to detect initial edge locations before using this algorithm. The templates of OFMM are improved, the condition have changed when we judge which pixel is edge. (2) A novel algorithm of automatic reference point detection is put forward and implemented. Firstly, the Canny operator is utilized to detect initial edge locations in pixel. Secondly, useless edges are filtered by using inherent attributes of the ellipse, such as length, roundness, convex-concave and enclosure. Thirdly, accurate edges in sub-pixel are obtained by using the improved orthogonal Fourier-Mellin moment operator. Fourthly, reference points are classified by using the improved method. At last, sub-pixel center coordinates are identified. (3) A matching algorithm based on camera poses is put forward. Firstly, poses of the camera are obtained based on coded targets which are matched. Secondly, uncoded targets are matched by using epipolar constraint based on camera poses obtained. At last, refined matching is performed based on initial matching by means of symmetrical distance constraint to increase number of corresponding points and to improve the accuracy of the matching between corresponding points.更多还原