基于特征点和方向向量的快速目标匹配算法研究

发布时间：2018-01-18 12:31

本文关键词：基于特征点和方向向量的快速目标匹配算法研究　出处：《上海交通大学》2015年硕士论文　论文类型：学位论文

【摘要】：随着视觉传感器的快速发展和各类计算机视觉算法的成熟,越来越多的机器视觉检测方法运用到了工业生产检测领域。视觉检测方法具有非接触性、高速度、高精度和高度自动化等特点,能够使人们远离危险的工作场合,并且极大地提高生产效率,适合于流水线上大批量目标物的定位与检测。传统的目标匹配算法通常只兼顾匹配精度、速度中的某一方面。本文提出一种分层匹配的方法,首先基于ORB特征点,选取Hamming距离作为相似性度量标准,进行模板和搜索图之间的特征点匹配,并采用随机抽样一致算法去除大部分误匹配点,推算出最终的仿射变换参数。然后,利用匹配对中特征点的旋转角和模板相对位置,计算出待匹配图像中特征点对应的标记中心点,采用K-means聚类算法将这些标记中心点进行聚类,将分属于不同目标物副本的特征点分类。最后,采用改进的方向向量提取算法,比较模板和搜索图像的边缘信息,对目标位置进行精定位。为了进一步加速计算,文章采用了图像金字塔算法,从粗到精逐步确定模板在图像上的位置和旋转角度;为了使匹配精度达到亚像素等级,文章采用了多项式拟合的算法,通过对拟合出的三阶曲面求局部极大值,求得模板在图像中的亚像素位置和旋转角度。实验表明,本文提出的方法能够分离出图像中多个目标物副本,具有匹配速度快、准确性高、旋转不变性和光照不变性等特点。
[Abstract]:With the rapid development of the visual sensor and various computer vision algorithms, machine vision detection method is more and more applied to the field of industrial production testing. Visual detection method has the advantages of non-contact, high speed, high precision and high automation, the workplace can make people from danger, and greatly improve the production efficiency. Suitable for mass detection and orientation of target line. The traditional object matching algorithm is usually considered the matching accuracy, a speed. This paper proposes a hierarchical matching method, based on the ORB feature points, select the Hamming distance as the similarity measure, the feature points between the template and the search graph matching, and random sample consensus algorithm to remove most of the false matching points, calculate the affine transformation parameters finally. Then, using the matching of feature points The relative position of the rotation angle and the template, calculate the mark center point matching feature points in the image, the center point clustering of these markers using K-means clustering algorithm, the feature points belonging to different copies of the target classification. Finally, the improved direction vector extraction algorithm, compare the template and search the edge information of the image. Precise positioning of the target position. In order to further accelerate the computation, the algorithm of image Pyramid, from coarse to fine gradually determine position of the template in the image and rotation angle; in order to make the matching precision sub-pixel level, this paper adopts the polynomial fitting algorithm based on three order surface fitting for local maxima and obtain sub-pixel position template in the image and rotation angle. Experimental results show that the proposed method can separate multiple copies of the target image, with matching speed Fast, high accuracy, rotation invariance and illumination invariance and so on.

【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41

【参考文献】