OpenCV的SIFT特征点提取及RANSAC去除误检测_opencv sift ransac-优快云博客

本文链接：https://blog.youkuaiyun.com/sunzhao1000/article/details/80590174

关于SIFT的讲解，参考：http://blog.youkuaiyun.com/zddblog/article/details/7521424

关于SANCAS的讲解，参考：http://blog.youkuaiyun.com/laobai1015/article/details/51682596

SIFT（Scale Invariant Feature Transform）多尺度特征点提取算法在nonfree文件内集成。

RANSAC（Random Sample Consensus）随机抽样一致算法在calib3d文件内集成。

主要的思路是：

（1）利用SiftFeatureDetector类对象提取SIFT特征点；

（2）利用SiftDescriptorExtractor类对象进行特征点的特征提取；

（3）利用BFMatcher类对象实现初步匹配；

（注意：在第3步中，实际上已经完成SIFT特征点的提取与匹配，但是会有较多的无匹配，因此后面的步骤主要思路是利用RANSAC方法去除误匹配）

（4）利用2-范数特征阈值初步筛选去除误匹配；

（6）利用findFundamentalMat函数实现RANSAC算法。

文件头如下：

#include <opencv2/opencv.hpp>
#include <iostream>
#include <opencv2/nonfree/nonfree.hpp>
#include <opencv2/features2d/features2d.hpp>

代码如下：

Mat Limg = imread("Limage.png", CV_LOAD_IMAGE_ANYDEPTH);
Mat Rimg = imread("Rimage.png", CV_LOAD_IMAGE_ANYDEPTH);
// detecting keypoints
SiftFeatureDetector detector(0, 3, 0.04, 10, 1.6);
// (特征点数目-按sift特征质量排序，金字塔中每组的层数，过滤的阈值，边缘效应阈值，高斯滤波系数)
vector<KeyPoint> Lkeypoints, Rkeypoints;
detector.detect(Limg, Lkeypoints);
detector.detect(Rimg, Rkeypoints);
// computing descriptors
// SiftDescriptorExtractor extractor;
SiftDescriptorExtractor extractor;
Mat Ldescriptors, Rdescriptors;
extractor.compute(Limg, Lkeypoints, Ldescriptors);
extractor.compute(Rimg, Rkeypoints, Rdescriptors);
// matching descriptors
BFMatcher matcher(NORM_L1); //> 对应sift
vector<DMatch> matches;
matcher.match(Ldescriptors, Rdescriptors, matches);
//
namedWindow("matches", CV_WINDOW_FREERATIO);
Mat img_matches;
drawMatches(Limg, Lkeypoints, Rimg, Rkeypoints, matches, img_matches);
imshow("matches", img_matches);
// coordination of match-feature-point
vector<Point2f> Lp, Rp;
for (size_t m = 0; m < matches.size(); m++)
{
int i1 = matches[m].queryIdx;
int i2 = matches[m].trainIdx;
Mat Ldesc = Ldescriptors.row(i1);
Mat Rdesc = Rdescriptors.row(i2);
double EuclideanDist = norm(Ldesc, Rdesc);
if (EuclideanDist <= 50)
{
const KeyPoint &kp1 = Lkeypoints[i1], &kp2 = Rkeypoints[i2];
Lp.push_back(kp1.pt);
Rp.push_back(kp2.pt);
}
}
// 用RANSAC方法计算基本矩阵&去除误匹配
Mat Fundamental;
vector<uchar> RANSACStatus;
Fundamental = findFundamentalMat(Lp, Rp, RANSACStatus, CV_FM_RANSAC, 3.0, 0.995);
// 筛选特征点
vector<Point2f> Lpoint, Rpoint;
vector<DMatch> m_match;
int idxnum = 0;
for (int i = 0; i < RANSACStatus.size(); i++)
{
if (RANSACStatus[i] != 0)
{
Lpoint.push_back(Lp[i]);
Rpoint.push_back(Rp[i]);
idxnum++;
}
}
m_match.resize(idxnum);
for (int idx = 0; idx < idxnum; idx++)
{
// 用于匹配验证
m_match[idx].queryIdx = idx;
m_match[idx].trainIdx = idx;
}
// 把内点转换为drawMatches可以使用的格式
vector<KeyPoint> Lkp(idxnum);
vector<KeyPoint> Rkp(idxnum);
KeyPoint::convert(Lpoint, Lkp);
KeyPoint::convert(Rpoint, Rkp);
// 显示筛选后的特征点
namedWindow("m_match", CV_WINDOW_FREERATIO);
Mat m_matches;
drawMatches(Limg, Lkp, Rimg, Rkp, m_match, m_matches);
imshow("m_match", m_matches);