OpenCV4中的ORB特征提取与匹配

微信公众号：OpenCV学堂

关注获取更多计算机视觉与深度学习知识

FAST特征点定位

假设存在连续N个点满足

BRIEF特征描述子

旋转不变性

其中n=p+q称为N阶距。得到中心坐标为：

计算得到角度为：

OpenCV中基于ORB特征对象匹配代码演示如下：

#include 
#include 
#include 
#define RATIO    0.4
using namespace cv;
using namespace std;
int main(int argc, char** argv) {
    Mat box = imread("D:/vcprojects/images/box.png");
    Mat scene = imread("D:/vcprojects/images/box_in_scene.png");
    if (scene.empty()) {
        printf("could not load image...\n");
        return -1;
    }
    imshow("input image", scene);
    vector<KeyPoint> keypoints_obj, keypoints_sence;
    Mat descriptors_box, descriptors_sence;
    Ptr<ORB> detector = ORB::create();
    detector->detectAndCompute(scene, Mat(), keypoints_sence, descriptors_sence);
    detector->detectAndCompute(box, Mat(), keypoints_obj, descriptors_box);
    vector<DMatch> matches;
    // 初始化flann匹配
    // Ptr matcher = FlannBasedMatcher::create(); // default is bad, using local sensitive hash(LSH)
    Ptr<DescriptorMatcher> matcher = makePtr<FlannBasedMatcher>(makePtr<flann::LshIndexParams>(12, 20, 2));
    matcher->match(descriptors_box, descriptors_sence, matches);
    // 发现匹配
    vector<DMatch> goodMatches;
    printf("total match points : %d\n", matches.size());
    float maxdist = 0;
    for (unsigned int i = 0; i < matches.size(); ++i) {
        printf("dist : %.2f \n", matches[i].distance);
        maxdist = max(maxdist, matches[i].distance);
    }
    for (unsigned int i = 0; i < matches.size(); ++i) {
        if(matches[i].distance < maxdist*RATIO)
            goodMatches.push_back(matches[i]);
    }
    Mat dst;
    drawMatches(box, keypoints_obj, scene, keypoints_sence, goodMatches, dst);
    imshow("output", dst);
    waitKey(0);
    return 0;
}

在OpenCV中基于BRIEF与ORB特征实现FLANN对象匹配经常会遇到一个错误如下：

Ptr<DescriptorMatcher> matcher = makePtr<FlannBasedMatcher>
                          (makePtr<flann::LshIndexParams>(12, 20, 2));

运行结果：

为学患无疑，疑则进也