OpenCV 之 角点检测
角点 (corners) 的定义有两个版本:一是 两条边缘的交点,二是 邻域内具有两个主方向的特征点。
一般而言,角点是边缘曲线上曲率为极大值的点,或者 图像亮度发生剧烈变化的点。例如,从人眼角度来看,下图的 $E$ 和 $F$ 便是典型的角点
1 检测思路
在图像中定义一个局部小窗口,然后沿各个方向移动这个窗口,则会出现 a) b) c) 三种情况,分别对应平坦区、边缘和角点
a) 窗口内的图像强度,在窗口向各个方向移动时,都没有发生变化,则窗口内都是 “平坦区”,不存在角点
b) 窗口内的图像强度,在窗口向某一个 (些) 方向移动时,发生较大变化;而在另一些方向不发生变化,那么,窗口内可能存在 “边缘”
c) 窗口内的图像强度,在窗口向各个方向移动时,都发生了较大的变化,则认为窗口内存在 “角点”
a) flat region b) edge c) corner
2 Harris 角点
2.1 公式推导
图像在点 $(x,y) $ 处的灰度值为 $I(x, y)$,当在 $x$ 方向上平移 $u$,且 $y$ 方向上平移 $v$ 时,图像灰度值的变化为
$ \qquad E(u,v) = \sum\limits_{x,y} \, \underbrace{w(x,y)}_\text{window function} \; [\underbrace{I(x+u, y+v)}_\text{shifted intensity} – \underbrace{I(x, y)}_\text{intensity}]^2 $
一阶泰勒级数近似展开得
$ \qquad \sum\limits_{x,y} \; [I(x+u, y+v) – I(x, y)]^2 \approx \sum\limits_{x,y} \; [I(x, y) +uI_x + vI_y – I(x, y)]^2 = \sum\limits_{x,y} \; [u^2I_x^2 + 2uvI_x I_y + v^2I_y^2 ] $
写成矩阵形式
$ \qquad E(u,v) \approx \begin{bmatrix} u & v \end{bmatrix} \left ( \displaystyle \sum_{x,y} w(x,y) \begin{bmatrix} I_x^{2} & I_{x}I_{y} \\ I_xI_{y} & I_{y}^{2} \end{bmatrix} \right ) \begin{bmatrix} u \\ v \end{bmatrix}$
则有
$ \qquad E(u,v) \approx \begin{bmatrix} u & v \end{bmatrix} M \begin{bmatrix} u \\ v \end{bmatrix}$, 假定 $ M = \displaystyle \sum_{x,y} w(x,y) \begin{bmatrix} I_x^{2} & I_{x}I_{y} \\ I_xI_{y} & I_{y}^{2} \end{bmatrix}$
2.2 判别方法
定义一个角点响应值$\qquad R = det(M) – k(trace(M))^{2} = \lambda_{1} \lambda_{2} – k (\lambda_{1}+\lambda_{2})^2 $
根据响应值的大小,判断小窗口内是否包含角点:
1) “平坦区”:|R| 小的区域,即 $\lambda_1$ 和 $\lambda_2$ 都小;
2) “边缘”: R <0 的区域,即 $\lambda_1 >> \lambda_2$ 或反之;
3) “角点”: R 大的区域,即 $\lambda_1$ 和 $\lambda_2$ 都大且近似相等
为了便于直观理解,绘制成 $\lambda_1-\lambda_2$ 平面如下图:
2.3 cornerHarris()
OpenCV 中 Harris 角点检测的函数为:
void cv::cornerHarris ( InputArray src, // 输入图像 (单通道,8位或浮点型) OutputArray dst, // 输出图像 (类型 CV_32FC1,大小同 src) int blockSize, // 邻域大小 int ksize, // Sobel 算子的孔径大小 double k, // 经验参数,取值范围 0.04 ~ 0.06 int borderType = BORDER_DEFAULT // 边界模式 )
2.4 代码示例
#include "opencv2/highgui.hpp" #include "opencv2/imgproc.hpp" using namespace cv; // Harris corner parameters int kThresh = 150; int kBlockSize = 2; int kApertureSize = 3; double k = 0.04; int main() { // read image Mat src, src_gray; src = imread("building.jpg"); if(src.empty()) return -1; cvtColor(src, src_gray, COLOR_BGR2GRAY); Mat dst, dst_norm, dst_norm_scaled; // Harris corner detect cornerHarris(src_gray, dst, kBlockSize, kApertureSize, k); normalize(dst, dst_norm, 0, 255, NORM_MINMAX, CV_32FC1); convertScaleAbs(dst_norm, dst_norm_scaled); // draw detected corners for(int j=0; j < dst_norm.rows; j++) { for(int i=0; i<dst_norm.cols;i++) { if((int)dst_norm.at<float>(j,i) > kThresh) { circle(src, Point(i, j), 2, Scalar(0,255,0)); } } } imshow("harris corner", src); waitKey(0); }
检测结果:
3 Shi-Tomasi 角点
Shi-Tomasi 角点是 Harris 角点的改进,在多数情况下,其检测效果要优于 Harris。二者的区别在于,Shi-Tomasi 选取 $\lambda_1$ 和 $\lambda_2$ 中的最小值,作为新的角点响应值 $R$
$\qquad R = min(\lambda_1, \lambda_2) $
则相应的 $\lambda_1-\lambda_2$ 平面为:
3.1 goodFeaturesToTrack()
OpenCV 中 Shi-Tomasi 角点检测函数为:
void cv::goodFeaturesToTrack ( InputArray image, // 输入图像 (单通道,8位或浮点型32位) OutputArray corners, // 检测到的角点 int maxCorners, // 最多允许返回的角点数量 double qualityLevel, // double minDistance, // 角点间的最小欧拉距离 InputArray mask = noArray(), // int blockSize = 3, // bool useHarrisDetector = false, // double k = 0.04 // )
3.2 代码示例
#include "opencv2/highgui.hpp" #include "opencv2/imgproc.hpp" using namespace cv; using namespace std; int kMaxCorners = 1000; double kQualityLevel = 0.1; double kMinDistance = 1; int main() { // read image Mat src, src_gray; src = imread("building.jpg"); if (src.empty()) return -1; cvtColor(src, src_gray, COLOR_BGR2GRAY); // Shi-Tomasi corner detect vector<Point2f> corners; goodFeaturesToTrack(src_gray, corners, kMaxCorners, kQualityLevel, kMinDistance); // draw and show detected corners for (size_t i = 0; i < corners.size(); i++) { circle(src, corners[i], 2.5, Scalar(0, 255, 0)); } imshow("Shi-Tomasi corner", src); waitKey(0); }
检测结果:
4 角点检测的实现
在 OpenCV 中分析 cornerHarris() 函数的源码,得到实现步骤如下:sobel 算子求解 dx 和 dy -> 矩阵 M -> boxFilter -> 每个像素的角点响应值 R
代码实现:
#include <iostream> #include "opencv2/highgui.hpp" #include "opencv2/imgproc.hpp" using namespace cv; using namespace std; int kApertureSize = 3; int kBlockSize = 2; double k = 0.04; int kThresh = 150; int main() { // read image Mat src, src_gray; src = imread("chessboard.png"); if (src.empty()) return -1; cvtColor(src, src_gray, COLOR_BGR2GRAY);
// scale int depth = src_gray.depth(); double scale = (double)(1 << (2* kBlockSize)); if (depth == CV_8U) scale *= 255.0; scale = 1.0 / scale; // 1) dx, dy Mat Dx, Dy; Sobel(src_gray, Dx, CV_32F, 1, 0, kApertureSize, scale); Sobel(src_gray, Dy, CV_32F, 0, 1, kApertureSize, scale); // 2) cov Size size = src_gray.size(); Mat cov(size, CV_32FC3); for (int i = 0; i < size.height; i++) { float* cov_data = cov.ptr<float>(i); const float* dxdata = Dx.ptr<float>(i); const float* dydata = Dy.ptr<float>(i); for (int j=0; j < size.width; j++) { float dx = dxdata[j]; float dy = dydata[j]; cov_data[j * 3] = dx * dx; cov_data[j * 3 + 1] = dx * dy; cov_data[j * 3 + 2] = dy * dy; } } // 3) boxfilter boxFilter(cov, cov, cov.depth(), Size(kBlockSize, kBlockSize)); // 4) R Mat dst(size,CV_32FC1); Size size_cov = cov.size(); for (int i = 0; i < size_cov.height; i++) { const float* ptr_cov = cov.ptr<float>(i); float* ptr_dst = dst.ptr<float>(i); for (int j=0; j < size_cov.width; j++) { float a = ptr_cov[j * 3]; float b = ptr_cov[j * 3 + 1]; float c = ptr_cov[j * 3 + 2]; ptr_dst[j] = (float)(a * c - b * b - k * (a + c) * (a + c)); } } // 5) normalize Mat dst_norm, dst_norm_scaled; normalize(dst, dst_norm, 0, 255, NORM_MINMAX, CV_32FC1); convertScaleAbs(dst_norm, dst_norm_scaled); // 6) draw detected corners for (int j = 0; j < dst_norm.rows; j++) { for (int i = 0; i < dst_norm.cols; i++) { if ((int)dst_norm.at<float>(j, i) > 150) { circle(src, Point(i, j), 2, Scalar(0, 255, 0)); } } } imshow("Harris corner", src); waitKey(0); }
5 亚像素角点检测
亚像素角点的提取函数 cornerSubPix(),常用于相机标定中,定义如下:
5.1 cornerSubpix()
void cv::cornerSubPix( InputArray image, // 输入图象(单通道,8位或浮点型) InputOutputArray corners, // 亚像素精度的角点坐标 Size winSize, // 搜索窗口尺寸的 1/2 Size zeroZone, // TermCriteria criteria // 迭代终止准则 )
5.2 代码示例
#include <iostream> #include "opencv2/highgui.hpp" #include "opencv2/imgproc.hpp" using namespace cv; using namespace std; int kMaxCorners = 40; double kQualityLevel = 0.01; double kMinDistance = 50; int main() { // 1) read image Mat src, src_gray; src = imread("chessboard.png"); if (src.empty()) return -1; cvtColor(src, src_gray, COLOR_BGR2GRAY); // 2) Shi-Tomasi corner detect vector<Point2f> corners; goodFeaturesToTrack(src_gray, corners, kMaxCorners, kQualityLevel, kMinDistance); // 3) draw and show detected corners for (size_t i = 0; i < corners.size(); i++) { circle(src, corners[i], 3, Scalar(0, 255, 0)); } imshow("Shi-Tomasi corner", src); TermCriteria criteria = TermCriteria(TermCriteria::EPS + TermCriteria::COUNT, 40, 0.001); // 4) find corner positions in subpixel cornerSubPix(src_gray, corners, Size(5, 5), Size(-1, -1), criteria); // 5) output subpixel corners for (size_t i = 0; i < corners.size(); i++) { cout << "Corner[" << i << "]: (" << corners[i].x << "," << corners[i].y << ")" << endl; } waitKey(0); }
输入棋盘格5行8列,对应7×4个角点,图像的分辨率为 600*387,则所有角点的理论坐标如下表:
角点的图象坐标值输出如下:
参考资料:
《图像局部不变性特征与描述》 第 3 章
https://www.cnblogs.com/ronny/p/4009425.html
http://www.cse.psu.edu/~rtc12/CSE486/
OpenCV Tutorials / feature2d module / Harris corner detector
OpenCV-Python Tutorials / Feature Detection and Description / Shi-Tomasi Corner Detector & Good Features to Track
OpenCV Tutorials / feature2d module / Creating your own corner detector
OpenCV Tutorials / feature2d module / Detecting corners location in subpixels