基于ransac算法的sift特征匹配研究(OpenCV+VS2010)____朱万革(最终版)

SHANGHAI JIAO TONG UNIVERSITY

学士学位论文

THESIS OF BACHELOR

基于ransac算法的sift特征匹配研究（OpenCV+VS2010）

上海交通大学

毕业设计（论文）学术诚信声明

本人郑重声明：所呈交的毕业设计（论文），是本人在导师的指导下，独立进行研究工作所取得的成果。除文中已经注明引用的内容外，本论文不包含任何其他个人或集体已经发表或撰写过的作品成果。对本文的研究做出重要贡献的个人和集体，均已在文中以明确方式标明。本人完全意识到本声明的法律结果由本人承担。

作者签名：

日期：年月日

上海交通大学

毕业设计（论文）版权使用授权书

本毕业设计（论文）作者同意学校保留并向国家有关部门或机构送交论文的复印件和电子版，允许论文被查阅和借阅。本人授权上海交通大学可以将本毕业设计（论文）的全部或部分内容编入有关数据库进行检索，可以采用影印、缩印或扫描等复制手段保存和汇编本毕业设计（论文）。

保密□，在年解密后适用本授权书。

本论文属于

不保密□。

（请在以上方框内打“√”）

作者签名：指导教师签名：

日期：年月日日期：年月日

视频图像跟踪系统

摘要

图像（Image）--是客观世界的景物通过光学系统作用后产生的影像。图像直观地反映了场景中物体的颜色、亮度等特征，从而使我们能清晰分辨他们的形状、大小和空间位置。近30年来人们试图研究基于计算机的视觉系统，并且试图利用其系统来代替工业农业上的有害劳动。这样的视觉系统渐渐地进入我们的生活，让我们的生活变得很丰富，并且我们现在享受着图像处理这学问的成果。在世界上的先进的国家都设立了图像处理研究所，研究解决国防部门所要的问题。

本文将介绍基于OpenCV（Open Source Computer Vision Library）的视频图像匹配、拼接、融合和目标跟踪的算法以及方法。说到图像拼接，本文中所用的图像拼接算法是高效的SIFT特征算法。首先，用两个通用的USB摄像头来实时地进行采集图像，并对这两幅图像提取SIFT特征点。然后，进行粗匹配。最后用RANSAC算法对所提取出来的SIFT特征点匹配对进行提纯以及估计模型参数。最后把两幅图像拼接成一幅完整的图像，并且用加权平均算法进行无缝拼接。再进行摄像头标定，求出两个通用摄像头的内外参数，最后进行测距以及跟踪。最终取得了令人满意的结果。

关键词：SIFT，匹配,拼接，配准，RANSAC

VIDEO TRACKING SYSTEM

ABSTRACT

For many years, people have been studying how to make the robot or the computer able to identify targets and obtain information about the surrounding environment. We people can easily see and identify every kind of objects, but for computers or robots, this is a very difficult task and it is a process that involves a lot of scientific knowledge. The main part of object recognition is digital image processing. After the invention of the computer, people began to direct their research on how to make the computer more powerful and useful. For this purpose, many scientists have dedicated their life for the development of computer. The rapid development of computer causes a very fast development of digital image processing. Why we people study science? Of course the answer will be to make our life easier, and to be able to live in our dream life, so that we can enjoy the life in comfort and happiness.

Nowadays, Image processing technology is everywhere around us, but sometimes because we are used to this technolog y so we don’t pay attention. For example, the phone's handwriting input method, company entrance fingerprint identification system, license plate recognition system, robotics system program for exploring the lunar, medical imaging technology, facial recognition systems, and satellite imaging system and so on. In the last three decades, image processing technology has made a rapid development, which is inseparable from the development of computers, and more inseparable from the development of materials science. We can notice that science now have penetrated into every professional image processing and the image comes to many areas. These days image processing technology is directly related to our standard life, this technology involves image recognition, image analysis and image stitching, etc. Image processing is now facing enormous challenges, due to the development of materials processing industry, CNC machine tools and control theory, image processing technology requirements are very high, Therefore, many scientists have spent their life studying image processing technology, trying to develop more flexible, more reliable, more accurate image processing technology and image processing algorithms.

Video tracking system includes the image stitching technology, when we mention image stitching technology, we have to talk about image matching and image registration, because these two modules are the core parts of the image stitching.

In this study, I used Scale-invariant feature transform (SIFT) algorithm, this algorithm features repeatability, unique, localized, quantitative, accuracy and efficiency.

First from the two cameras (people left and right eye) in synchronous reads the image sequence, and I applied these image sequences RANSAC algorithm based on SIFT feature matching and obtained a good image stitching. Then this image with an image template matching of image recognition and tracking (based on SIFT), while supporting the binocular measurements to obtain distance information. Typically, about two a moment to read the two camera images with a lot SIFT feature point, so the need to purify the data using the RANSAC algorithm, like that