Sub pixel corner detection opencv cornersubpix

Opencv-cornerSubPix Principle introduction

If the purpose of image processing is not to extract feature points for recognition, but to carry out geometric measurement , This usually requires higher accuracy , The function goodFeaturesToTrack() Only simple pixel coordinate values can be provided , That is to say, sometimes we need real coordinate values instead of integer coordinate values .

Corner location features ： The connecting line between the corner and the edge point is perpendicular to the gradient direction of the edge point .

As shown in the figure above , Suppose a starting corner q Near the actual sub-pixel corner .

p The point is q In the neighborhood near the point , if p The point is inside the uniform area , be p The gradient of the point is 0; if p Point on the edge , be p The gradient direction of the point is perpendicular to the edge direction . If the vector q-p The direction is consistent with the edge direction , that q-p Vector and p The result of dot product operation of gradient vector of point is 0 .

At the initial corner ( The initial corner may not be on the edge ) Nearby, we can collect the gradients and correlation vectors of many groups of points q-p , At this time q It is the more accurate corner position we require , Then the vector dot product of each group is set to 0 , It is based on this idea , Product points into 0 The equations of form a system equation , The solution of the system equation is a more accurate sub-pixel corner position . New q Point as the center of the region , You can continue to use this method for iteration , Obtain high accuracy .

Look for sub-pixel corners ：cornerSubPix function

void cornerSubPix(InputArray image, InputOutputArray corners, Size winSize, Size zeroZone, TermCriteria criteria);

image, The input image , Source image .
corners, Provide the initial coordinates of the input corner and the precise output coordinates .
winSize, Half the size of the search window . if winSize = Size(5,5), That means using （52+1）×（52+1）= 11×11 Size search window .
zeroZone, Half the size of the dead zone . The real search area is [zeroZone2+1 , winSize2+1]. The value is （-1,-1） Indicates that there is no dead zone .
criteria, The termination condition of the iterative process of finding the corner . Or the number of iterations is greater than a certain set value , Or the accuracy reaches a certain set value , It can even be a combination of them .

#include<opencv.hpp>
#include<iostream>
#include<string>
#include<vector>
using namespace std;
using namespace cv;
Mat src, cornerImg, grayImg;
int mxCorners = 10;
RNG rngs = {
     12345 };
void Change(int, void*) {
    
    // Calculate the integer corner 
    vector<Point>corners1;
    vector<Point2f>corners2;    //https://www.cnblogs.com/bjxqmy/p/12459005.html
    goodFeaturesToTrack(grayImg, corners1, mxCorners, 0.01, 10, Mat(), 3, false, 0.04);
    goodFeaturesToTrack(grayImg, corners2, mxCorners, 0.01, 10, Mat(), 3, false, 0.04);
    
    // Calculate the exact corner 
    TermCriteria criteria = TermCriteria(TermCriteria::EPS + TermCriteria::MAX_ITER, 40, 0.001);
    cornerSubPix(grayImg, corners2, Size(5, 5), Size(-1, -1), criteria);

    Mat dst = src.clone();
    for (int i = 0; i < corners1.size(); i++) {
    
        cout << "[" << corners1[i].x << "," << corners1[i].y << "]" << endl;
        cout << "[" << corners2[i].x << "," << corners2[i].y << "]" << endl << endl;

        Scalar colors = Scalar(rngs.uniform(0, 255), rngs.uniform(0, 255), rngs.uniform(0, 255));
        circle(dst, corners2[i], 5, colors, -1);
    }
    imshow("dst", dst);
}
int main() {
    
    src = imread("C:/Users/ Qi Mingyang /Desktop/ The sample picture /1.jpg");
    imshow("src", src);

    // Convert to grayscale image 
    cvtColor(src, grayImg, COLOR_BGR2GRAY);

    namedWindow("dst");
    createTrackbar("maxCorners", "dst", &mxCorners, 100, Change);
    Change(0, 0);

    waitKey(0);
}

Reference source ：https://blog.csdn.net/weixin_41695564/article/details/79991733