OpenCV 中有兩種特徵匹配方法:暴力匹配 (Brute force matching) 和 最近鄰匹配 (Nearest Neighbors matching)html
它們都繼承自 DescriptorMatcher,是基於特徵描述符距離的匹配,根據描述符的不一樣,距離能夠是 歐氏距離,也能夠是 漢明距算法
首先,任取圖像 A 的一個特徵描述符,計算它到圖像 B 中全部特徵描述符的距離;而後,將所獲得的距離進行排序;最後,選擇距離最短的特徵,做爲 A-B 的匹配點app
BFMatcher 屬於 features2d 模塊,繼承自 DescriptorMatcher,其 create() 函數以下: 函數
static Ptr<BFMatcher> create( int normType = NORM_L2, // normType, One of NORM_L1, NORM_L2, NORM_HAMMING, NORM_HAMMING2. bool crossCheck = false // crossCheck );
1) normType 距離類型ui
SIFT和SURF 的 HOG 描述符,對應歐氏距離 L1 和 L2;ORB 和 BRISK 的 BRIEF 描述符,對應漢明距 HAMMING;HAMMING2 則對應當 WTA_K = 3或4 時的 ORB 算法this
- 歐氏距離:最經常使用的一種距離定義,指的是 n 維空間中,兩點之間的實際距離spa
$L1 = \sum_I | \texttt{src1} (I) - \texttt{src2}|$3d
$L2 = \sqrt{\sum_I (\texttt{src1}(I) - \texttt{src2}(I))^2}$ rest
- 漢明距離:實際是計算機的異或操做,適用於二進制串描述符,如 BRIEF 描述符,定義以下:code
$ Hamming \left ( a,b \right ) = \sum\limits_{i=0}^{n-1} \left ( a_i \oplus b_i \right ) $
2) crossCheck 交叉覈對
- 若是在圖像 B 中,特徵 $f_{b}$ 是特徵 $f_{a}$ 的最佳匹配,而且在圖像 A 中,特徵 $f_{a}$ 也是特徵 $f_{b}$ 的最佳匹配,則稱 $(f_{a}, f_{b})$ 爲 "good match"
特徵匹配步驟以下:讀圖 -> 提取特徵 -> 計算特徵描述符 -> 暴力匹配 -> 顯示匹配結果
#include "opencv2/highgui.hpp" #include "opencv2/features2d.hpp" using namespace cv; int main() { // 1) read Mat img1 = imread("box.png", IMREAD_GRAYSCALE); Mat img2 = imread("box_in_scene.png", IMREAD_GRAYSCALE); if (img1.empty() || img2.empty()) return -1; // 2) detect and compute Ptr<SIFT> sift = SIFT::create(); std::vector<KeyPoint> kps1, kps2; Mat desc1, desc2; sift->detectAndCompute(img1, Mat(), kps1, desc1); sift->detectAndCompute(img2, Mat(), kps2, desc2); // 3) match Ptr<BFMatcher> bfmatcher = BFMatcher::create(NORM_L2, true); std::vector<DMatch> matches; bfmatcher->match(desc1, desc2, matches); // 4) draw and show Mat img_matches; drawMatches(img1, kps1, img2, kps2, matches, img_matches); imshow("BFMatcher", img_matches); waitKey(); }
crosscheck 分別爲 true 和 false:
FLANN 是一個開源庫,全稱 Fast Library for Approximate Nearest Neighbors,它實現了一系列高維向量的近似最近鄰搜索算法
基於 FLANN 庫的最近鄰匹配算子 FlannBasedMatcher,在特徵數據集較大或一些實時處理領域,其運行效率要遠高於 BFMatcher
OpenCV 中 FlannBasedMatcher 的定義以下:
// This matcher trains cv::flann::Index on a train descriptor collection and calls its nearest search methods to find the best matches.
// So, this matcher may be faster when matching a large train collection than the brute force matcher. class FlannBasedMatcher : public DescriptorMatcher { public: FlannBasedMatcher( const Ptr<flann::IndexParams>& indexParams=makePtr<flann::KDTreeIndexParams>(), const Ptr<flann::SearchParams>& searchParams=makePtr<flann::SearchParams>() ); static Ptr<FlannBasedMatcher> create();
爲了進一步提升特徵匹配精度,David Lowe 提出了一種最近鄰次近鄰距離比的方法:
- 取圖像 A 的一個特徵,搜索它到圖像 B 距離最近的兩個特徵,距離分別記爲 $d_{1}$ 和 $d_{2}$,只有當 $\displaystyle{\frac{d_{1}}{d_{2}}}$ 小於某個閾值時,才認爲是 "good match"
"good match" 的機率密度函數 PDF (Probability Density Function) 與最近鄰次近鄰距離比的關係,以下:
取 distance ratio = 0.7,對比使用和不使用距離比濾波的匹配效果,代碼以下:
#include "opencv2/highgui.hpp" #include "opencv2/features2d.hpp" using namespace cv; const float kRatioThresh = 0.7f; int main() { // 1) read Mat img1 = imread("box.png", IMREAD_GRAYSCALE); Mat img2 = imread("box_in_scene.png", IMREAD_GRAYSCALE); if (img1.empty() || img2.empty()) return -1; // 2) detect feature and compute descriptor Ptr<SIFT> sift = SIFT::create(); std::vector<KeyPoint> kps1, kps2; Mat desc1, desc2; sift->detectAndCompute(img1, Mat(), kps1, desc1); sift->detectAndCompute(img2, Mat(), kps2, desc2); // 3) FLANN based matcher Ptr<FlannBasedMatcher> knnmatcher = FlannBasedMatcher::create(); std::vector<std::vector<DMatch> > matches; knnmatcher->knnMatch(desc1, desc2, matches, 2); // 4) filter matches using Lowe's distance ratio test std::vector<DMatch> good_matches; for (size_t i = 0; i < matches.size(); i++) { if (matches[i][0].distance < kRatioThresh*matches[i][1].distance) { good_matches.push_back(matches[i][0]); } } // 5) draw and show matches Mat img_matches; drawMatches(img1, kps1, img2, kps2, good_matches, img_matches); imshow("Good Matches", img_matches);
waitKey(); }
匹配效果對好比下:
特徵匹配 + 平面單應性,在計算機視覺中有不少應用,如:透視校訂,目標定位等
OpenCV 之 平面單應性 4.1 中的示例,並非標準的透視校訂,由於是人拿着標定板旋轉不一樣角度,使相機和標定板產生了相對的視角變換,而不是相機和整個場景之間
多視圖幾何中,嚴格意思的透視校訂,是指相機在不一樣的視角下,對同一場景成不一樣的像而進行的視角校訂,以下圖:
在獲得匹配點對 good_matches 以後,再執行以下代碼,即可用於透視校訂
// Localize the object std::vector<Point2f> obj; std::vector<Point2f> scene; for (size_t i = 0; i < good_matches.size(); i++) { // Get the keypoints from the good matches obj.push_back(kps1[good_matches[i].queryIdx].pt); scene.push_back(kps2[good_matches[i].trainIdx].pt); } // estimate H Mat H = findHomography(scene, obj, RANSAC); // warp scene Mat scene_warp; warpPerspective(img2, scene_warp, H, Size(1.35*img2.cols, img2.rows)); // show imshow("scene_warp", scene_warp);
校訂先後的結果以下:
獲得匹配點對 good_matches 後,再執行以下代碼,即可用於目標定位
// Localize the object std::vector<Point2f> obj; std::vector<Point2f> scene; for (size_t i = 0; i < good_matches.size(); i++) { // Get the keypoints from the good matches obj.push_back(kps1[good_matches[i].queryIdx].pt); scene.push_back(kps2[good_matches[i].trainIdx].pt); } // estimate H Mat H = findHomography(obj,scene, RANSAC); // get the corners from the image_1 ( the object to be "detected" ) std::vector<Point2f> obj_corners(4); obj_corners[0] = Point2f(0, 0); obj_corners[1] = Point2f((float)img1.cols, 0); obj_corners[2] = Point2f((float)img1.cols, (float)img1.rows); obj_corners[3] = Point2f(0, (float)img1.rows); std::vector<Point2f> scene_corners(4); perspectiveTransform(obj_corners, scene_corners, H); // draw lines between the corners (the mapped object in the scene - image_2 ) line(img_matches, scene_corners[0] + Point2f((float)img1.cols, 0), scene_corners[1] + Point2f((float)img1.cols, 0), Scalar(0,255,0)); line(img_matches, scene_corners[1] + Point2f((float)img1.cols, 0), scene_corners[2] + Point2f((float)img1.cols, 0), Scalar(0,255,0)); line(img_matches, scene_corners[2] + Point2f((float)img1.cols, 0), scene_corners[3] + Point2f((float)img1.cols, 0), Scalar(0,255,0)); line(img_matches, scene_corners[3] + Point2f((float)img1.cols, 0), scene_corners[0] + Point2f((float)img1.cols, 0), Scalar(0,255,0)); // show detected matches imshow("Object detection", img_matches);
目標定位結果以下:
OpenCV-Python Tutorials / Feature Detection and Description / Feature Matching
OpenCV Tutorials / 2D Features framework (feature2d module) / Feature Matching with FLANN
OpenCV Tutorials / 2D Features framework (feature2d module) / Features2D + Homography to find a known object