SIFT特征匹配技术研究与应用
发布时间：2018-08-27 05:52
【摘要】：图像已经成为信息化时代下人们获取信息的一种必要手段,如何利用图像处理技术获取外界信息成为国内外研究者重点关注的一类问题。尺度不变特征变换(Scale Invariant Feature Transform,SIFT)算法因其在图像尺度变化、旋转等状况下的鲁棒性和独特性在特征匹配中得到了广泛的应用,然而该算法在特征点生成时效性和匹配精度上仍有一定的局限性。本文针对计算机图像处理中目标识别和目标跟踪两大研究方向,引入经典的SIFT算法的思想并对其进行优化,设计了改进的目标匹配和运动目标跟踪算法。本论文的主要研究内容包括:(1)使用体现图像信息量的图像熵进行关键点阈值判断,设计了自适应的关键点阈值调整方法;(2)引入基于直方图距离计算的EMD距离,同时基于SIFT算法特性,将改进EMD算法与多梯度方向SIFT特征点相结合进行距离的比对和运算的剪枝;(3)针对于多目标识别,设计了基于SIFT特征点双向匹配的改进算法;(4)设计一种融合SIFT向量和DBSCAN聚类的方法,以替代TLD算法中的跟踪模块。且对TLD算法检测模块进行调整。根据上述设计思路,本文实现了基于改进的SIFT算法的目标识别和目标跟踪算法,并通过测试数据集对所设计的算法进行了验证。实验结果表明本文方法能够(1)较好的解决图像匹配中多数特征点无意义匹配的问题;(2)较好的解决了匹配过程中诸多场景下欧氏距离不适用的问题;(3)实现多目标场景中的识别检测;(4)较好的解决TLD算法的跟踪模块在运动目标长期跟踪中难以保持鲁棒跟踪的问题。本文所设计的方法,对经典的SIFT算法的不足之处做出了针对性的改进,不仅提高了图像目标的的匹配准确度,并且运算效率相对于原算法也有了较好的进步,在目标的识别和跟踪应用中具有更好的适用性。
[Abstract]:Image has become a necessary means for people to obtain information in the information age. How to use image processing technology to obtain external information has become the focus of attention of researchers at home and abroad. Scale invariant feature transform (Scale Invariant Feature Transform,SIFT) algorithm is widely used in feature matching because of its robustness and uniqueness in image scale change and rotation. However, the algorithm still has some limitations in feature point generation timeliness and matching accuracy. Aiming at the two research directions of target recognition and target tracking in computer image processing, this paper introduces the idea of classical SIFT algorithm and optimizes it, and designs an improved target matching and moving target tracking algorithm. The main research contents of this thesis are as follows: (1) using image entropy to judge the threshold of key points, and designing an adaptive threshold adjustment method for key points; (2) introducing the EMD distance based on histogram distance. At the same time, based on the characteristics of SIFT algorithm, the improved EMD algorithm is combined with multi-gradient direction SIFT feature points to carry out distance comparison and pruning. (3) for multi-target recognition, An improved algorithm based on bidirectional matching of SIFT feature points is designed. (4) A method of combining SIFT vector and DBSCAN clustering is designed to replace the tracking module in TLD algorithm. And the TLD algorithm detection module is adjusted. According to the above design ideas, this paper implements the target recognition and target tracking algorithm based on the improved SIFT algorithm, and verifies the designed algorithm through the test data set. The experimental results show that this method can (1) solve the problem of meaningless matching of most feature points in image matching, (2) solve the problem that Euclidean distance is not suitable for many scenes in the matching process, and (3) realize the multi-objective field. (4) solve the problem that the tracking module of TLD algorithm is difficult to keep robust tracking in the long term tracking of moving target. The method designed in this paper improves the shortcomings of the classical SIFT algorithm. It not only improves the accuracy of image target matching, but also improves the computational efficiency compared with the original algorithm. It has better applicability in target recognition and tracking applications.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2206297
[Abstract]:Image has become a necessary means for people to obtain information in the information age. How to use image processing technology to obtain external information has become the focus of attention of researchers at home and abroad. Scale invariant feature transform (Scale Invariant Feature Transform,SIFT) algorithm is widely used in feature matching because of its robustness and uniqueness in image scale change and rotation. However, the algorithm still has some limitations in feature point generation timeliness and matching accuracy. Aiming at the two research directions of target recognition and target tracking in computer image processing, this paper introduces the idea of classical SIFT algorithm and optimizes it, and designs an improved target matching and moving target tracking algorithm. The main research contents of this thesis are as follows: (1) using image entropy to judge the threshold of key points, and designing an adaptive threshold adjustment method for key points; (2) introducing the EMD distance based on histogram distance. At the same time, based on the characteristics of SIFT algorithm, the improved EMD algorithm is combined with multi-gradient direction SIFT feature points to carry out distance comparison and pruning. (3) for multi-target recognition, An improved algorithm based on bidirectional matching of SIFT feature points is designed. (4) A method of combining SIFT vector and DBSCAN clustering is designed to replace the tracking module in TLD algorithm. And the TLD algorithm detection module is adjusted. According to the above design ideas, this paper implements the target recognition and target tracking algorithm based on the improved SIFT algorithm, and verifies the designed algorithm through the test data set. The experimental results show that this method can (1) solve the problem of meaningless matching of most feature points in image matching, (2) solve the problem that Euclidean distance is not suitable for many scenes in the matching process, and (3) realize the multi-objective field. (4) solve the problem that the tracking module of TLD algorithm is difficult to keep robust tracking in the long term tracking of moving target. The method designed in this paper improves the shortcomings of the classical SIFT algorithm. It not only improves the accuracy of image target matching, but also improves the computational efficiency compared with the original algorithm. It has better applicability in target recognition and tracking applications.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前5条
1 李宗艳;宋丽梅;习江涛;郭庆华;朱新军;陈明磊;;A stereo matching algorithm based on SIFT feature and homography matrix[J];Optoelectronics Letters;2015年05期
2 华莉琴;许维;王拓;马瑞芳;胥博;;采用改进的尺度不变特征转换及多视角模型对车型识别[J];西安交通大学学报;2013年04期
3 曾峦;王元钦;谭久彬;;改进的SIFT特征提取和匹配算法[J];光学精密工程;2011年06期
4 王梅;屠大维;周许超;;SIFT特征匹配和差分相乘融合的运动目标检测[J];光学精密工程;2011年04期
5 于丽莉;戴青;;一种改进的SIFT特征匹配算法[J];计算机工程;2011年02期
，本文编号：2206297
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2206297.html
