如何分析ORB、SURF、SIFT特征点提取方法以及ICP匹配方法

2023-06-22 05:06:33 628人浏览八月长安

摘要

本篇文章给大家分享的是有关如何分析ORB、SURF、SIFT特征点提取方法以及ICP匹配方法，小编觉得挺实用的，因此分享给大家学习，希望大家阅读完这篇文章后可以有所收获，话不多说，跟着小编一起来看看吧。在进行编译视觉SLAM时，书中提到了O

本篇文章给大家分享的是有关如何分析ORB、SURF、SIFT特征点提取方法以及ICP匹配方法，小编觉得挺实用的，因此分享给大家学习，希望大家阅读完这篇文章后可以有所收获，话不多说，跟着小编一起来看看吧。

在进行编译视觉SLAM时，书中提到了ORB、SURF、SIFT提取方法，以及特征提取方法暴力匹配(Brute-Force Matcher)和快速近邻匹配(FLANN)。以及7.9讲述的3D-3D：迭代最近点(Iterative Closest Point,ICP)方法，ICP 的求解方式有两种：利用线性代数求解(主要是SVD)，以及利用非线性优化方式求解。

main.cpp

#include <iOStream> #include "OpenCV2/opencv.hpp"#include "opencv2/core/core.hpp"#include "opencv2/features2d/features2d.hpp"#include "opencv2/highgui/highgui.hpp"#include <opencv2/xfeatures2d.hpp>#include <iostream>#include <vector>#include <time.h>#include <chrono>#include <math.h>#include<bits/stdc++.h>  using namespace std;using namespace cv;using namespace cv::xfeatures2d; double picture1_size_change=1;double picture2_size_change=1; bool show_picture = true; void extract_ORB2(string picture1, string picture2){    //-- 读取图像    Mat img_1 = imread(picture1, CV_LOAD_IMAGE_COLOR);    Mat img_2 = imread(picture2, CV_LOAD_IMAGE_COLOR);    assert(img_1.data != nullptr && img_2.data != nullptr);    resize(img_1, img_1, Size(),  picture1_size_change, picture1_size_change);    resize(img_2, img_2, Size(), picture2_size_change, picture2_size_change);     //-- 初始化    std::vector<KeyPoint> keypoints_1, keypoints_2;    Mat descriptors_1, descriptors_2;    Ptr<FeatureDetector> detector = ORB::create(2000,(1.200000048F), 8, 100);    Ptr<DescriptorExtractor> descriptor = ORB::create(5000);    Ptr<DescriptORMatcher> matcher = DescriptorMatcher::create("BruteForce-Hamming");     //-- 第一步:检测 Oriented FAST 角点位置    chrono::steady_clock::time_point t1 = chrono::steady_clock::now();    detector->detect(img_1, keypoints_1);    detector->detect(img_2, keypoints_2);     //-- 第二步:根据角点位置计算 BRIEF 描述子    descriptor->compute(img_1, keypoints_1, descriptors_1);    descriptor->compute(img_2, keypoints_2, descriptors_2);    chrono::steady_clock::time_point t2 = chrono::steady_clock::now();    chrono::duration<double> time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);    // cout << "extract ORB cost = " << time_used.count() * 1000 << " ms " << endl;    cout << "detect " << keypoints_1.size() << " and " << keypoints_2.size() << " keypoints " << endl;     if (show_picture)    {        Mat outimg1;        drawKeypoints(img_1, keypoints_1, outimg1, Scalar::all(-1), DrawMatchesFlags::DEFAULT);        imshow("ORB features", outimg1);    }     //-- 第三步:对两幅图像中的BRIEF描述子进行匹配，使用 Hamming 距离    vector<DMatch> matches;    // t1 = chrono::steady_clock::now();    matcher->match(descriptors_1, descriptors_2, matches);    t2 = chrono::steady_clock::now();    time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);    cout << "extract and match ORB cost = " << time_used.count() * 1000 << " ms " << endl;      //-- 第四步:匹配点对筛选    // 计算最小距离和最大距离    auto min_max = minmax_element(matches.begin(), matches.end(),                                  [](const DMatch &m1, const DMatch &m2)                                  { return m1.distance < m2.distance; });    double min_dist = min_max.first->distance;    double max_dist = min_max.second->distance;     //   printf("-- Max dist : %f \n", max_dist);    //   printf("-- Min dist : %f \n", min_dist);     //当描述子之间的距离大于两倍的最小距离时,即认为匹配有误.但有时候最小距离会非常小,设置一个经验值30作为下限.    std::vector<DMatch> Good_matches;    for (int i = 0; i < descriptors_1.rows; i++)    {        if (matches[i].distance <= max(2 * min_dist, 30.0))        {            good_matches.push_back(matches[i]);        }    }        cout << "match " << good_matches.size() << " keypoints " << endl;      //-- 第五步:绘制匹配结果    Mat img_match;    Mat img_goodmatch;    drawMatches(img_1, keypoints_1, img_2, keypoints_2, matches, img_match);    drawMatches(img_1, keypoints_1, img_2, keypoints_2, good_matches, img_goodmatch);     if (show_picture)        imshow("good matches", img_goodmatch);    if (show_picture)        waiTKEy(0);} void extract_SIFT(string picture1, string picture2){    // double t = (double)getTickCount();    Mat temp = imread(picture1, IMREAD_GRAYSCALE);    Mat image_check_changed = imread(picture2, IMREAD_GRAYSCALE);    if (!temp.data || !image_check_changed.data)    {        printf("could not load images...\n");        return;    }     resize(temp, temp, Size(),  picture1_size_change, picture1_size_change);    resize(image_check_changed, image_check_changed, Size(), picture2_size_change, picture2_size_change);     //Mat image_check_changed = Change_image(image_check);    //("temp", temp);    if (show_picture)         imshow("image_check_changed", image_check_changed);     int minHessian = 500;    // Ptr<SURF> detector = SURF::create(minHessian);    // surf    Ptr<SIFT> detector = SIFT::create(); // sift     vector<KeyPoint> keypoints_obj;    vector<KeyPoint> keypoints_scene;    Mat descriptor_obj, descriptor_scene;      clock_t startTime, endTime;    startTime = clock();     chrono::steady_clock::time_point t1 = chrono::steady_clock::now();    // cout << "extract ORB cost = " << time_used.count() * 1000 << " ms " << endl;    detector->detectAndCompute(temp, Mat(), keypoints_obj, descriptor_obj);    detector->detectAndCompute(image_check_changed, Mat(), keypoints_scene, descriptor_scene);    cout << "detect " << keypoints_obj.size() << " and " << keypoints_scene.size() << " keypoints " << endl;     // matching    FlannBasedMatcher matcher;    vector<DMatch> matches;    matcher.match(descriptor_obj, descriptor_scene, matches);     chrono::steady_clock::time_point t2 = chrono::steady_clock::now();    chrono::duration<double> time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);    cout << "extract and match cost = " << time_used.count() * 1000 << " ms " << endl;     //求最小最大距离    double minDist = 1000;    double maxDist = 0;    //row--行 col--列    for (int i = 0; i < descriptor_obj.rows; i++)    {        double dist = matches[i].distance;        if (dist > maxDist)        {            maxDist = dist;        }        if (dist < minDist)        {            minDist = dist;        }    }    // printf("max distance : %f\n", maxDist);    // printf("min distance : %f\n", minDist);     // find good matched points    vector<DMatch> goodMatches;    for (int i = 0; i < descriptor_obj.rows; i++)    {        double dist = matches[i].distance;        if (dist < max(5 * minDist, 1.0))        {            goodMatches.push_back(matches[i]);        }    }    //rectangle(temp, Point(1, 1), Point(177, 157), Scalar(0, 0, 255), 8, 0);     cout << "match " << goodMatches.size() << " keypoints " << endl;     endTime = clock();    // cout << "took time : " << (double)(endTime - startTime) / CLOCKS_PER_SEC * 1000 << " ms" << endl;     Mat matchesImg;    drawMatches(temp, keypoints_obj, image_check_changed, keypoints_scene, goodMatches, matchesImg, Scalar::all(-1),                Scalar::all(-1), vector<char>(), DrawMatchesFlags::NOT_DRAW_SINGLE_POINTS);    if (show_picture)         imshow("Flann Matching Result01", matchesImg);    // imwrite("C:/Users/Administrator/Desktop/matchesImg04.jpg", matchesImg);     //求h    std::vector<Point2f> points1, points2;     //保存对应点    for (size_t i = 0; i < goodMatches.size(); i++)    {        //queryIdx是对齐图像的描述子和特征点的下标。        points1.push_back(keypoints_obj[goodMatches[i].queryIdx].pt);        //queryIdx是是样本图像的描述子和特征点的下标。        points2.push_back(keypoints_scene[goodMatches[i].trainIdx].pt);    }     // Find homography 计算Homography，RANSAC随机抽样一致性算法    Mat H = findHomography(points1, points2, RANSAC);    //imwrite("C:/Users/Administrator/Desktop/C-train/C-train/result/sift/Image4_SURF_MinHessian1000_ minDist1000_a0.9b70.jpg", matchesImg);     vector<Point2f> obj_corners(4);    vector<Point2f> scene_corners(4);    obj_corners[0] = Point(0, 0);    obj_corners[1] = Point(temp.cols, 0);    obj_corners[2] = Point(temp.cols, temp.rows);    obj_corners[3] = Point(0, temp.rows);     //透视变换(把斜的图片扶正)    perspectiveTransform(obj_corners, scene_corners, H);    //Mat dst;    cvtColor(image_check_changed, image_check_changed, COLOR_GRAY2BGR);    line(image_check_changed, scene_corners[0], scene_corners[1], Scalar(0, 0, 255), 2, 8, 0);    line(image_check_changed, scene_corners[1], scene_corners[2], Scalar(0, 0, 255), 2, 8, 0);    line(image_check_changed, scene_corners[2], scene_corners[3], Scalar(0, 0, 255), 2, 8, 0);    line(image_check_changed, scene_corners[3], scene_corners[0], Scalar(0, 0, 255), 2, 8, 0);      if (show_picture)    {        Mat outimg1;        Mat temp_color = imread(picture1, CV_LOAD_IMAGE_COLOR);        drawKeypoints(temp_color, keypoints_obj, outimg1, Scalar::all(-1), DrawMatchesFlags::DEFAULT);        imshow("SIFT features", outimg1);    }     if (show_picture)         imshow("Draw object", image_check_changed);    // imwrite("C:/Users/Administrator/Desktop/image04.jpg", image_check_changed);     // t = ((double)getTickCount() - t) / getTickFrequency();    // printf("averagetime:%f\n", t);    if (show_picture)         waitKey(0);} void extract_SURF(string picture1, string picture2){     // double t = (double)getTickCount();    Mat temp = imread(picture1, IMREAD_GRAYSCALE);    Mat image_check_changed = imread(picture2, IMREAD_GRAYSCALE);    if (!temp.data || !image_check_changed.data)    {        printf("could not load images...\n");        return;    }     resize(temp, temp, Size(),  picture1_size_change, picture1_size_change);    resize(image_check_changed, image_check_changed, Size(), picture2_size_change, picture2_size_change);     //Mat image_check_changed = Change_image(image_check);    //("temp", temp);    if (show_picture)         imshow("image_check_changed", image_check_changed);     int minHessian = 500;    Ptr<SURF> detector = SURF::create(minHessian);    // surf    // Ptr<SIFT> detector = SIFT::create(minHessian); // sift     vector<KeyPoint> keypoints_obj;    vector<KeyPoint> keypoints_scene;    Mat descriptor_obj, descriptor_scene;      clock_t startTime, endTime;    startTime = clock();     chrono::steady_clock::time_point t1 = chrono::steady_clock::now();    // cout << "extract ORB cost = " << time_used.count() * 1000 << " ms " << endl;    detector->detectAndCompute(temp, Mat(), keypoints_obj, descriptor_obj);    detector->detectAndCompute(image_check_changed, Mat(), keypoints_scene, descriptor_scene);    cout << "detect " << keypoints_obj.size() << " and " << keypoints_scene.size() << " keypoints " << endl;     // matching    FlannBasedMatcher matcher;    vector<DMatch> matches;    matcher.match(descriptor_obj, descriptor_scene, matches);     chrono::steady_clock::time_point t2 = chrono::steady_clock::now();    chrono::duration<double> time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);    cout << "extract and match cost = " << time_used.count() * 1000 << " ms " << endl;     //求最小最大距离    double minDist = 1000;    double maxDist = 0;    //row--行 col--列    for (int i = 0; i < descriptor_obj.rows; i++)    {        double dist = matches[i].distance;        if (dist > maxDist)        {            maxDist = dist;        }        if (dist < minDist)        {            minDist = dist;        }    }    // printf("max distance : %f\n", maxDist);    // printf("min distance : %f\n", minDist);     // find good matched points    vector<DMatch> goodMatches;    for (int i = 0; i < descriptor_obj.rows; i++)    {        double dist = matches[i].distance;        if (dist < max(2 * minDist, 0.15))        {            goodMatches.push_back(matches[i]);        }    }    //rectangle(temp, Point(1, 1), Point(177, 157), Scalar(0, 0, 255), 8, 0);     cout << "match " << goodMatches.size() << " keypoints " << endl;    endTime = clock();    // cout << "took time : " << (double)(endTime - startTime) / CLOCKS_PER_SEC * 1000 << " ms" << endl;     Mat matchesImg;    drawMatches(temp, keypoints_obj, image_check_changed, keypoints_scene, goodMatches, matchesImg, Scalar::all(-1),                Scalar::all(-1), vector<char>(), DrawMatchesFlags::NOT_DRAW_SINGLE_POINTS);    if (show_picture)         imshow("Flann Matching Result01", matchesImg);    // imwrite("C:/Users/Administrator/Desktop/matchesImg04.jpg", matchesImg);     //求h    std::vector<Point2f> points1, points2;     //保存对应点    for (size_t i = 0; i < goodMatches.size(); i++)    {        //queryIdx是对齐图像的描述子和特征点的下标。        points1.push_back(keypoints_obj[goodMatches[i].queryIdx].pt);        //queryIdx是是样本图像的描述子和特征点的下标。        points2.push_back(keypoints_scene[goodMatches[i].trainIdx].pt);    }     // Find homography 计算Homography，RANSAC随机抽样一致性算法    Mat H = findHomography(points1, points2, RANSAC);    //imwrite("C:/Users/Administrator/Desktop/C-train/C-train/result/sift/Image4_SURF_MinHessian1000_ minDist1000_a0.9b70.jpg", matchesImg);     vector<Point2f> obj_corners(4);    vector<Point2f> scene_corners(4);    obj_corners[0] = Point(0, 0);    obj_corners[1] = Point(temp.cols, 0);    obj_corners[2] = Point(temp.cols, temp.rows);    obj_corners[3] = Point(0, temp.rows);     //透视变换(把斜的图片扶正)    perspectiveTransform(obj_corners, scene_corners, H);    //Mat dst;    cvtColor(image_check_changed, image_check_changed, COLOR_GRAY2BGR);    line(image_check_changed, scene_corners[0], scene_corners[1], Scalar(0, 0, 255), 2, 8, 0);    line(image_check_changed, scene_corners[1], scene_corners[2], Scalar(0, 0, 255), 2, 8, 0);    line(image_check_changed, scene_corners[2], scene_corners[3], Scalar(0, 0, 255), 2, 8, 0);    line(image_check_changed, scene_corners[3], scene_corners[0], Scalar(0, 0, 255), 2, 8, 0);      if (show_picture)    {        Mat outimg1;        Mat temp_color = imread(picture1, CV_LOAD_IMAGE_COLOR);        drawKeypoints(temp_color, keypoints_obj, outimg1, Scalar::all(-1), DrawMatchesFlags::DEFAULT);        imshow("SURF features", outimg1);    }     if (show_picture)         imshow("Draw object", image_check_changed);    // imwrite("C:/Users/Administrator/Desktop/image04.jpg", image_check_changed);     // t = ((double)getTickCount() - t) / getTickFrequency();    // printf("averagetime:%f\n", t);    if (show_picture)         waitKey(0);}void extract_AKAZE(string picture1,string picture2){    //读取图片    Mat temp = imread(picture1,IMREAD_GRAYSCALE);    Mat image_check_changed = imread(picture2,IMREAD_GRAYSCALE);    //如果不能读到其中任何一张图片，则打印不能下载图片    if(!temp.data || !image_check_changed.data)    {        printf("could not load iamges...\n");        return;    }    resize(temp,temp,Size(),picture1_size_change,picture1_size_change);    resize(image_check_changed,image_check_changed,Size(),picture2_size_change,picture2_size_change);     //Mat image_check_changed = Change_image(image_check);    //("temp", temp);     if(show_picture)    {        imshow("image_checked_changed",image_check_changed);    }        int minHessian=500;    Ptr<AKAZE> detector=AKAZE::create();//AKAZE     vector<keypoint> keypoints_obj;    vector<keypoint> keypoints_scene;    Mat descriptor_obj,descriptor_scene;      clock_t startTime,endTime;    startTime=clock();     chrono::steady_clock::time_point t1=chrono::steady_clock::now();    detector->detectAndCompute(temp,Mat(),keypoints_obj,descriptor_obj);    detector->detectAndCompute(image_check_changed,Mat(),keypoints_scene,descriptor_scene);    cout<<" detect "<<keypoints_obj.size()<<" and "<<keypoints_scene.size<<" keypoints "<<endl;     //matching    FlannBasedMatcher matcher;    vector<DMatch> matches;    matcher.match(descriptor_obj,descriptor_scene,matches);     chrono::steady_clock::time_point t2 = chrono::steady_clock::now();    chrono::duration<double> time_used = chrono::duration_cast<chrono::duration<double>>(t2-t1);    cout << "extract and match cost = " << time_used.count()*1000<<" ms "<<endl;     //求最小最大距离    double minDist = 1000;    double max_dist = 0;    //row--行 col--列    for(int i=0;i<descriptor_obj.rows;i++)    {        double dist = match[i].distance;        if(dist > maxDist)        {            maxDist = dist;        }        if(dist<minDist)        {            minDist = dist;        }    }    // printf("max distance : %f\n", maxDist);    // printf("min distance : %f\n", minDist);     // find good matched points    vector<DMatch> goodMatches;    for(imt i=0;i<descriptor_obj.rows;i++)    {        double dist = matches[i].distance;        if(dist < max(5 * minDist,1.0))        {        goodMatches.push_back(matches[i]);        }    }    //rectangle(temp, Point(1, 1), Point(177, 157), Scalar(0, 0, 255), 8, 0);    cout<<" match "<<goodMatches.size()<<" keypoints "<<endl;    endTime = clock();    // cout << "took time : " << (double)(endTime - startTime) / CLOCKS_PER_SEC * 1000 << " ms" << endl;        Mat matchesImg;    drawMatches(temp,keypoints_obj,image_check_changed,keypoints_scene,goodMatches,    matchesImg,Scalar::all(-1),                   Scalar::all(-1),vector<char>(),DrawMatchesFlags::NOT_DRAW_SINGLE_POINTS);    if(show_picture)       imshow("Flann Matching Result01",matchesImg);        // imwrite("C:/Users/Administrator/Desktop/matchesImg04.jpg", matchesImg);     //求h       std::vector<Point2f> points1,points2;     //保存对应点    for(size_t i = 0;i < goodMatches.size();i++)    {       //queryIdx是对齐图像的描述子和特征点的下标。        points1.push_back(keypoints_obj[goodMatches[i].queryIdx].pt);        //queryIdx是是样本图像的描述子和特征点的下标。        points2.push_back(keypoints_scene[goodMatches[i].trainIdx].pt);      }    // Find homography 计算Homography，RANSAC随机抽样一致性算法    Mat H = findHomography(points1,points2,RANSAC);   //imwrite("C:/Users/Administrator/Desktop/C-train/C-train/result/sift/Image4_SURF_MinHessian1000_ minDist1000_a0.9b70.jpg", matchesImg);    vector<Point2f> obj_corners(4);   vector<Point2f> scene_corners(4);   obj_corners[0] = Point(0,0);   obj_corners[0] = Point(temp.count,0);   obj_corners[0] = Point(temp.cols,temp.rows);   obj_corners[0] = Point(0,temp.rows);    //透视变换(把斜的图片扶正)   perspectiveTransform(obj_corners,scene_corners,H);   //Mat dst   cvtColor(image_check_changed,image_check_changed,COLOR_GRAY2BGR);   line(image_check_changed,scene_corners[0],scene_corners[1],Scalar(0,0,255),2,8,0);   line(image_check_changed,scene_corners[1],scene_corners[2],Scalar(0,0,255),2,8,0);   line(image_check_changed,scene_corners[2],scene_corners[3],Scalar(0,0,255),2,8,0);   line(image_check_changed,scene_corners[3],scene_corners[0],Scalar(0,0,255),2,8,0);     if(show_picture)   {       Mat outimg1;       Mat temp_color = imread(picture1,CV_LOAD_IMAGE_COLOR);       drawKeypoints(temp_color,keypoints_obj,outimg1,Scalar::all(-1),DrawMatchesFlags::DEFAULT);       imshow("AKAZE features",outimg1);   }   if(show_picture)      waitKey(0);} void extract_ORB(string picture1, string picture2){    Mat img_1 = imread(picture1);Mat img_2 = imread(picture2);     resize(img_1, img_1, Size(), picture1_size_change, picture1_size_change);    resize(img_2, img_2, Size(), picture2_size_change, picture2_size_change); if (!img_1.data || !img_2.data){cout << "error reading images " << endl;return ;} vector<Point2f> recognized;vector<Point2f> scene; recognized.resize(1000);scene.resize(1000); Mat d_srcL, d_srcR; Mat img_matches, des_L, des_R;//ORB算法的目标必须是灰度图像cvtColor(img_1, d_srcL, COLOR_BGR2GRAY);//CPU版的ORB算法源码中自带对输入图像灰度化，此步可省略cvtColor(img_2, d_srcR, COLOR_BGR2GRAY); Ptr<ORB> d_orb = ORB::create(1500); Mat d_descriptorsL, d_descriptorsR, d_descriptorsL_32F, d_descriptorsR_32F; vector<KeyPoint> keyPoints_1, keyPoints_2; //设置关键点间的匹配方式为NORM_L2，更建议使用 FLANNBASED = 1, BRUTEFORCE = 2, BRUTEFORCE_L1 = 3, BRUTEFORCE_HAMMING = 4, BRUTEFORCE_HAMMINGLUT = 5, BRUTEFORCE_SL2 = 6 Ptr<DescriptorMatcher> d_matcher = DescriptorMatcher::create(NORM_L2); std::vector<DMatch> matches;//普通匹配std::vector<DMatch> good_matches;//通过keyPoint之间距离筛选匹配度高的匹配结果     clock_t startTime, endTime;    startTime = clock();       chrono::steady_clock::time_point t1 = chrono::steady_clock::now();d_orb -> detectAndCompute(d_srcL, Mat(), keyPoints_1, d_descriptorsL);d_orb -> detectAndCompute(d_srcR, Mat(), keyPoints_2, d_descriptorsR);    cout << "detect " << keyPoints_1.size() << " and " << keyPoints_2.size() << " keypoints " << endl;    // endTime = clock();    // cout << "took time : " << (double)(endTime - startTime) / CLOCKS_PER_SEC * 1000 << " ms" << endl;  d_matcher -> match(d_descriptorsL, d_descriptorsR, matches);//L、R表示左右两幅图像进行匹配     //计算匹配所需时间    chrono::steady_clock::time_point t2 = chrono::steady_clock::now();    chrono::duration<double> time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);    cout << "extract and match cost = " << time_used.count() * 1000 << " ms " << endl;  int sz = matches.size();double max_dist = 0; double min_dist = 100; for (int i = 0; i < sz; i++){double dist = matches[i].distance;if (dist < min_dist) min_dist = dist;if (dist > max_dist) max_dist = dist;}  for (int i = 0; i < sz; i++){if (matches[i].distance < 0.6*max_dist){good_matches.push_back(matches[i]);}}     cout << "match " << good_matches.size() << " keypoints " << endl;    // endTime = clock();    // cout << "took time : " << (double)(endTime - startTime) / CLOCKS_PER_SEC * 1000 << " ms" << endl;  //提取良好匹配结果中在待测图片上的点集，确定匹配的大概位置for (size_t i = 0; i < good_matches.size(); ++i){scene.push_back(keyPoints_2[ good_matches[i].trainIdx ].pt);} for(unsigned int j = 0; j < scene.size(); j++)cv::circle(img_2, scene[j], 2, cv::Scalar(0, 255, 0), 2);//画出普通匹配结果Mat ShowMatches;drawMatches(img_1,keyPoints_1,img_2,keyPoints_2,matches,ShowMatches);if (show_picture)         imshow("matches", ShowMatches);// imwrite("matches.png", ShowMatches);//画出良好匹配结果Mat ShowGoodMatches;drawMatches(img_1,keyPoints_1,img_2,keyPoints_2,good_matches,ShowGoodMatches);if (show_picture)         imshow("good_matches", ShowGoodMatches);// imwrite("good_matches.png", ShowGoodMatches);//画出良好匹配结果中在待测图片上的点集if (show_picture)         imshow("MatchPoints_in_img_2", img_2);// imwrite("MatchPoints_in_img_2.png", img_2);if (show_picture)         waitKey(0);} int main(int arGC, char **argv){    string picture1=string(argv[1]);    string picture2=string(argv[2]);    // string picture1 = "data/picture1/6.jpg";    // string picture2 = "data/picture2/16.PNG";     cout << "\nextract_ORB::" << endl;    extract_ORB(picture1, picture2);     cout << "\nextract_ORB::" << endl;    extract_ORB2(picture1, picture2);     cout << "\nextract_SURF::" << endl;    extract_SURF(picture1, picture2);       cout << "\nextract_AKAZE::" << endl;    extract_AKAZE(picture1, picture2);       cout << "\nextract_SIFT::" << endl;    extract_SIFT(picture1, picture2);    cout << "success!!" << endl;}

CMakeLists.txt

CMAKE_MINIMUM_REQUIRED(VERSION 2.8.3)  # 设定版本PROJECT(DescriptorCompare) # 设定工程名SET(CMAKE_CXX_COMPILER "g++")  # 设定编译器add_compile_options(-std=c++14)   #编译选项，选择c++版本 # 设定可执行二进制文件的目录（最后生成的可执行文件放置的目录）SET(EXECUTABLE_OUTPUT_PATH ${PROJECT_SOURCE_DIR}) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -fpermissive -g -O3  -Wno-unused-function  -Wno-return-type") find_package(OpenCV 3.0 REQUIRED)  message(STATUS "Using opencv version ${OpenCV_VERSION}")find_package(Eigen3 3.3.8 REQUIRED)find_package(Pangolin REQUIRED)     # 设定链接目录LINK_DIRECTORIES(${PROJECT_SOURCE_DIR}/lib) # 设定头文件目录INCLUDE_DIRECTORIES(    ${PROJECT_SOURCE_DIR}/include    ${EIGEN3_INCLUDE_DIR}    ${OpenCV_INCLUDE_DIR}    ${Pangolin_INCLUDE_DIRS}    )     add_library(${PROJECT_NAME}test.cc) target_link_libraries( ${PROJECT_NAME}${OpenCV_LIBS}${EIGEN3_LIBS}${Pangolin_LIBRARIES} )  add_executable(main main.cpp )target_link_libraries(main ${PROJECT_NAME} )  add_executable(icp icp.cpp )target_link_libraries(icp ${PROJECT_NAME} )

执行效果

./main 1.png 2.png

extract_ORB::detect 1500 and 1500 keypoints extract and match cost = 21.5506 ms match 903 keypoints  extract_ORB::detect 1304 and 1301 keypoints extract and match ORB cost = 25.4976 ms match 313 keypoints  extract_SURF::detect 915 and 940 keypoints extract and match cost = 53.8371 ms match 255 keypoints  extract_SIFT::detect 1536 and 1433 keypoints extract and match cost = 97.9322 ms match 213 keypoints success!!

如何分析ORB、SURF、SIFT特征点提取方法以及ICP匹配方法

ICP

#include <iostream>#include <opencv2/core/core.hpp>#include <opencv2/features2d/features2d.hpp>#include <opencv2/highgui/highgui.hpp>#include <opencv2/calib3d/calib3d.hpp>#include <Eigen/Core>#include <Eigen/Dense>#include <Eigen/Geometry>#include <Eigen/SVD> #include <pangolin/pangolin.h>#include <chrono>  using namespace std;using namespace cv; int picture_h=480;int picture_w=640;  bool show_picture = true;  void find_feature_matches(  const Mat &img_1, const Mat &img_2,  std::vector<KeyPoint> &keypoints_1,  std::vector<KeyPoint> &keypoints_2,  std::vector<DMatch> &matches); // 像素坐标转相机归一化坐标Point2d pixel2cam(const Point2d &p, const Mat &K); void pose_estimation_3d3d(  const vector<Point3f> &pts1,  const vector<Point3f> &pts2,  Mat &R, Mat &t);   int main(int argc, char **argv) {  if (argc != 5) {    cout << "usage: pose_estimation_3d3d img1 img2 depth2 depth3" << endl;    return 1;  }  //-- 读取图像  Mat img_1 = imread(argv[1], CV_LOAD_IMAGE_COLOR);  Mat img_2 = imread(argv[2], CV_LOAD_IMAGE_COLOR);   vector<KeyPoint> keypoints_1, keypoints_2;  vector<DMatch> matches;  find_feature_matches(img_1, img_2, keypoints_1, keypoints_2, matches);  cout << "picture1 keypoints: " << keypoints_1.size() << " \npicture2 keypoints: " << keypoints_2.size() << endl;  cout << "一共找到了 " << matches.size() << " 组匹配点" << endl;   // 建立3D点  Mat depth2 = imread(argv[3], CV_8UC1);       // 深度图为16位无符号数，单通道图像  Mat depth3 = imread(argv[4], CV_8UC1);       // 深度图为16位无符号数，单通道图像  Mat K = (Mat_<double>(3, 3) << 595.2, 0, 328.9, 0, 599.0, 253.9, 0, 0, 1);  vector<Point3f> pts1, pts2;   for (DMatch m:matches) {    int d1 = 255-(int)depth2.ptr<uchar>(int(keypoints_1[m.queryIdx].pt.y))[int(keypoints_1[m.queryIdx].pt.x)];    int d2 = 255-(int)depth3.ptr<uchar>(int(keypoints_2[m.trainIdx].pt.y))[int(keypoints_2[m.trainIdx].pt.x)];    if (d1 == 0 || d2 == 0)   // bad depth      continue;    Point2d p1 = pixel2cam(keypoints_1[m.queryIdx].pt, K);    Point2d p2 = pixel2cam(keypoints_2[m.trainIdx].pt, K);    float dd1 = int(d1) / 1000.0;    float dd2 = int(d2) / 1000.0;    pts1.push_back(Point3f(p1.x * dd1, p1.y * dd1, dd1));    pts2.push_back(Point3f(p2.x * dd2, p2.y * dd2, dd2));  }   cout << "3d-3d pairs: " << pts1.size() << endl;  Mat R, t;  pose_estimation_3d3d(pts1, pts2, R, t);     //DZQ ADD  cv::Mat Pose = (Mat_<double>(4, 4) << R.at<double>(0, 0), R.at<double>(0, 1), R.at<double>(0, 2), t.at<double>(0),                  R.at<double>(1, 0), R.at<double>(1, 1), R.at<double>(1, 2), t.at<double>(1),                  R.at<double>(2, 0), R.at<double>(2, 1), R.at<double>(2, 2), t.at<double>(2),                  0, 0, 0, 1);   cout << "[delete outliers] Matched objects distance: ";  vector<double> vDistance;  double allDistance = 0; //存储总距离，用来求平均匹配距离，用平均的误差距离来剔除外点  for (int i = 0; i < pts1.size(); i++)  {    Mat point = Pose * (Mat_<double>(4, 1) << pts2[i].x, pts2[i].y, pts2[i].z, 1);    double distance = pow(pow(pts1[i].x - point.at<double>(0), 2) + pow(pts1[i].y - point.at<double>(1), 2) + pow(pts1[i].z - point.at<double>(2), 2), 0.5);    vDistance.push_back(distance);    allDistance += distance;    // cout << distance << " ";  }  // cout << endl;  double avgDistance = allDistance / pts1.size(); //求一个平均距离  int N_outliers = 0;  for (int i = 0, j = 0; i < pts1.size(); i++, j++) //i用来记录剔除后vector遍历的位置，j用来记录原位置  {    if (vDistance[i] > 1.5 * avgDistance) //匹配物体超过平均距离的N倍就会被剔除 [delete outliers]  DZQ FIXED_PARAM    {      N_outliers++;    }  }  cout << "N_outliers:: " << N_outliers << endl;        // show points  {    //创建一个窗口    pangolin::CreateWindowAndBind("show points", 640, 480);    //启动深度测试    glEnable(GL_DEPTH_TEST);     // Define Projection and initial ModelView matrix    pangolin::OpenGlRenderState s_cam(        pangolin::ProjectionMatrix(640, 480, 420, 420, 320, 240, 0.05, 500),        //对应的是gluLookAt,摄像机位置,参考点位置,up vector(上向量)        pangolin::ModelViewLookAt(0, -5, 0.1, 0, 0, 0, pangolin::AxisY));     // Create Interactive View in window    pangolin::Handler3D handler(s_cam);    //setBounds 跟opengl的viewport 有关    //看SimpleDisplay中边界的设置就知道    pangolin::View &d_cam = pangolin::CreateDisplay()                                .SetBounds(0.0, 1.0, 0.0, 1.0, -640.0f / 480.0f)                                .SetHandler(&handler);     while (!pangolin::ShouldQuit())    {       // Clear screen and activate view to render into      glClearColor(0.97,0.97,1.0, 1); //背景色       glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);      d_cam.Activate(s_cam);        glBegin(GL_POINTS);  //绘制匹配点      glLineWidth(5);      for (int i = 0; i < pts1.size(); i++)      {        glColor3f(1, 0, 0);        glVertex3d(pts1[i].x,pts1[i].y,pts1[i].z);        Mat point = Pose * (Mat_<double>(4, 1) << pts2[i].x, pts2[i].y, pts2[i].z, 1);        glColor3f(0, 1, 0);        glVertex3d(point.at<double>(0),point.at<double>(1),point.at<double>(2));      }      glEnd();        glBegin(GL_LINES);    //绘制匹配线      glLineWidth(1);      for (int i = 0; i < pts1.size(); i++)      {        glColor3f(0, 0, 1);        glVertex3d(pts1[i].x,pts1[i].y,pts1[i].z);        Mat point = Pose * (Mat_<double>(4, 1) << pts2[i].x, pts2[i].y, pts2[i].z, 1);        glVertex3d(point.at<double>(0),point.at<double>(1),point.at<double>(2));      }      glEnd();          glBegin(GL_POINTS);    //绘制所有点      glLineWidth(5);      glColor3f(1, 0.5, 0);      for (int i = 0; i < picture_h; i+=2)      {        for (int j = 0; j < picture_w; j+=2)        {          int d1 = 255-(int)depth2.ptr<uchar>(i)[j];          if (d1 == 0) // bad depth            continue;          Point2d temp_p;          temp_p.y=i;    //这里的x和y应该和i j相反          temp_p.x=j;          Point2d p1 = pixel2cam(temp_p, K);          float dd1 = int(d1) / 1000.0;          glVertex3d(p1.x * dd1, p1.y * dd1, dd1);          // glVertex3d(j/1000.0, i/1000.0, d1/200.0);        }      }      glEnd();         // Swap frames and Process Events      pangolin::FinishFrame();    }  }} void find_feature_matches(const Mat &img_1, const Mat &img_2,                          std::vector<KeyPoint> &keypoints_1,                          std::vector<KeyPoint> &keypoints_2,                          std::vector<DMatch> &matches) {  //-- 初始化  Mat descriptors_1, descriptors_2;  // used in OpenCV3  Ptr<FeatureDetector> detector = ORB::create(2000,(1.200000048F), 8, 100);  Ptr<DescriptorExtractor> descriptor = ORB::create(5000);    Ptr<DescriptorMatcher> matcher = DescriptorMatcher::create("BruteForce-Hamming");  //-- 第一步:检测 Oriented FAST 角点位置  detector->detect(img_1, keypoints_1);  detector->detect(img_2, keypoints_2);   //-- 第二步:根据角点位置计算 BRIEF 描述子  descriptor->compute(img_1, keypoints_1, descriptors_1);  descriptor->compute(img_2, keypoints_2, descriptors_2);   //-- 第三步:对两幅图像中的BRIEF描述子进行匹配，使用 Hamming 距离  vector<DMatch> match;  // BFMatcher matcher ( NORM_HAMMING );  matcher->match(descriptors_1, descriptors_2, match);   //-- 第四步:匹配点对筛选  double min_dist = 10000, max_dist = 0;   //找出所有匹配之间的最小距离和最大距离, 即是最相似的和最不相似的两组点之间的距离  for (int i = 0; i < descriptors_1.rows; i++) {    double dist = match[i].distance;    if (dist < min_dist) min_dist = dist;    if (dist > max_dist) max_dist = dist;  }   printf("-- Max dist : %f \n", max_dist);  printf("-- Min dist : %f \n", min_dist);   //当描述子之间的距离大于两倍的最小距离时,即认为匹配有误.但有时候最小距离会非常小,设置一个经验值30作为下限.  for (int i = 0; i < descriptors_1.rows; i++) {    if (match[i].distance <= max(2 * min_dist, 30.0)) {      matches.push_back(match[i]);    }  }   //-- 第五步:绘制匹配结果  if(show_picture)  {     Mat img_match;    Mat img_goodmatch;    drawMatches(img_1, keypoints_1, img_2, keypoints_2, matches, img_match);    imshow("all matches", img_match);    waitKey(0);  }} Point2d pixel2cam(const Point2d &p, const Mat &K) {  return Point2d(    (p.x - K.at<double>(0, 2)) / K.at<double>(0, 0),    (p.y - K.at<double>(1, 2)) / K.at<double>(1, 1)  );} void pose_estimation_3d3d(const vector<Point3f> &pts1,                          const vector<Point3f> &pts2,                          Mat &R, Mat &t) {  Point3f p1, p2;     // center of mass  int N = pts1.size();  for (int i = 0; i < N; i++) {    p1 += pts1[i];    p2 += pts2[i];  }  p1 = Point3f(Vec3f(p1) / N);  p2 = Point3f(Vec3f(p2) / N);  vector<Point3f> q1(N), q2(N); // remove the center  for (int i = 0; i < N; i++) {    q1[i] = pts1[i] - p1;    q2[i] = pts2[i] - p2;  }   // compute q1*q2^T  Eigen::Matrix3d W = Eigen::Matrix3d::Zero();  for (int i = 0; i < N; i++) {    W += Eigen::Vector3d(q1[i].x, q1[i].y, q1[i].z) * Eigen::Vector3d(q2[i].x, q2[i].y, q2[i].z).transpose();  }  // cout << "W=" << W << endl;   // SVD on W  Eigen::JacobiSVD<Eigen::Matrix3d> svd(W, Eigen::ComputeFullU | Eigen::ComputeFullV);  Eigen::Matrix3d U = svd.matrixU();  Eigen::Matrix3d V = svd.matrixV();    Eigen::Matrix3d R_ = U * (V.transpose());  if (R_.determinant() < 0) {    R_ = -R_;  }  Eigen::Vector3d t_ = Eigen::Vector3d(p1.x, p1.y, p1.z) - R_ * Eigen::Vector3d(p2.x, p2.y, p2.z);   // convert to cv::Mat  R = (Mat_<double>(3, 3) <<    R_(0, 0), R_(0, 1), R_(0, 2),    R_(1, 0), R_(1, 1), R_(1, 2),    R_(2, 0), R_(2, 1), R_(2, 2)  );  t = (Mat_<double>(3, 1) << t_(0, 0), t_(1, 0), t_(2, 0));}     void convertRGB2Gray(string picture){  double min;double max;  Mat depth_new_1 = imread(picture);       // 深度图为16位无符号数，单通道图像Mat test=Mat(20,256,CV_8UC3);int s;for (int i = 0; i < 20; i++) {  std::cout<<i<<" "; Vec3b* p = test.ptr<Vec3b>(i);for (s = 0; s < 32; s++) {p[s][0] = 128 + 4 * s;p[s][1] = 0;p[s][2] = 0;}p[32][0] = 255;p[32][1] = 0;p[32][2] = 0;for (s = 0; s < 63; s++) {p[33+s][0] = 255;p[33+s][1] = 4+4*s;p[33+s][2] = 0;}p[96][0] = 254;p[96][1] = 255;p[96][2] = 2;for (s = 0; s < 62; s++) {p[97 + s][0] = 250 - 4 * s;p[97 + s][1] = 255;p[97 + s][2] = 6+4*s;}p[159][0] = 1;p[159][1] = 255;p[159][2] = 254;for (s = 0; s < 64; s++) {p[160 + s][0] = 0;p[160 + s][1] = 252 - (s * 4);p[160 + s][2] = 255;}for (s = 0; s < 32; s++) {p[224 + s][0] = 0;p[224 + s][1] = 0;p[224 + s][2] = 252-4*s;}}   cout<<"depth_new_1 :: "<<depth_new_1.cols<<" "<<depth_new_1.rows<<" "<<endl;  Mat img_g=Mat(picture_h,picture_w,CV_8UC1);for(int i=0;i<picture_h;i++){   Vec3b *p = test.ptr<Vec3b>(0);  Vec3b *q = depth_new_1.ptr<Vec3b>(i);  for (int j = 0; j < picture_w; j++)  {        for(int k=0;k<256;k++)    {      if ( (((int)p[k][0] - (int)q[j][0] < 4) && ((int)q[j][0] - (int)p[k][0] < 4))&&           (((int)p[k][1] - (int)q[j][1] < 4) && ((int)q[j][1] - (int)p[k][1] < 4))&&           (((int)p[k][2] - (int)q[j][2] < 4) && ((int)q[j][2] - (int)p[k][2] < 4)))           {        img_g.at<uchar>(i,j)=k;      }    }  }} imwrite("14_Depth_3.png", img_g);waitKey();  }

CMakeLists.txt

和上面一样。

./icp 1.png 2.png 1_depth.png 2_depth.png

-- Max dist : 87.000000 -- Min dist : 4.000000 picture1 keypoints: 1304 picture2 keypoints: 1301一共找到了 313 组匹配点3d-3d pairs: 313[delete outliers] Matched objects distance: N_outliers:: 23