OpenCV学习笔记

cv::getRotationMatrix2D(center, angle_deg, scale) // 角度从图像上看为逆时针方向，从xy坐标看为顺时针方向

cv::String cv::format( const char* fmt, ... );

OpenCV Rect矩形框提取ROI，不能超出图像范围

cv::Mat image = cv::imread("xxx。jpg");
cv::Rect roi(-100,200, 500,500); // x< 0
cv::imshow("image(roi)", image(roi)); // image(roi)报错

cv::Mat image = cv::imread("test.jpg");
cv::Rect roi(100,200, image.rows+100, image.cols); // x< 0
cv::Mat img_show = image(roi); // image(roi)报错
cv::imshow("image(roi)", img_show);

OpenCV 矩形&|运算例程：

// Rect_(_Tp _x, _Tp _y, _Tp _width, _Tp _height);
cv::Rect r1(0, 0, 200, 100);
cv::Rect r2(250, 50, 200, 100);
cv::Rect rect_overlap = r1 & r2;
cv::Rect rect_sum = r1 | r2;

std::cout<< "r1 = " << r1 << std::endl;
std::cout<< "r2 = " << r2 << std::endl;
std::cout<< "rect_overlap = " << rect_overlap << std::endl;
std::cout<< "rect_sum     = " << rect_sum << std::endl;

运行结果如下：
r1 = [200 x 100 from (0, 0)]
r2 = [200 x 100 from (250, 50)]
rect_overlap = [0 x 0 from (0, 0)]
rect_sum     = [450 x 150 from (0, 0)]

测试cv::solvePnP,循环体最小耗时<0.4ms

    for (int k = 0; k < 100; k++)
    {

        const int64 time_start = cv::getTickCount();
        cv::Mat rvec, tvec;
        std::vector<cv::Point2f> image_points;
        std::vector<cv::Point3f> object_points;
        object_points.push_back(cv::Point3f(0.0f, 0.0f, 0.0f));
        object_points.push_back(cv::Point3f(1.0f, 0.0f, 0.0f));
        object_points.push_back(cv::Point3f(1.0f, 1.0f, 0.0f));
        object_points.push_back(cv::Point3f(0.0f, 1.0f, 0.0f));

        image_points.push_back(cv::Point2f(0.0f, 0.0f));
        image_points.push_back(cv::Point2f(3.0f, 0.0f));
        image_points.push_back(cv::Point2f(2.0f, 2.0f));
        image_points.push_back(cv::Point2f(0.0f, 3.0f));

        cv::solvePnP(
                    object_points, // 3-d points in object coordinate
                    image_points,  // 2-d points in image coordinates
                    intrinsic,     // Our camera matrix
                    distortion,    // distortion coefficients
                    rvec, // Output rotation *vector*.
                    tvec  // Output translation vector.
                    );

        printf("time span = %f secs
", (cv::getTickCount()-time_start)/cv::getTickFrequency());
    }

OpenCV多目标跟踪例程：

#include<tracking.hpp> 
#include<highgui.hpp> 
#include<video.hpp> 
#include<coreutility.hpp> 
#include<vector> 
using namespace cv; 
using namespace std; 
int main() 
{
    Mat frame; 
    VideoCapture cap("1.mp4");//输入待处理的视频 
    cap >> frame; 
    vector<Rect> rois; 
    selectROIs("rois", frame, rois, false); // GUI操作框选ROIs
    if (rois.size()<1) return 0;
    MultiTracker trackers;
    vector<Rect2d> obj;
    vector<Ptr<Tracker>> algorithms;
    for (auto i = 0; i < rois.size(); i++)
    {
        obj.push_back(rois[i]);
        algorithms.push_back(TrackerKCF::create());
    } 
    trackers.add(algorithms, frame, obj);
    while (cap.read(frame))
    {
        bool ok = trackers.update(frame);
        if (ok)
        {
            for (auto j = 0; j < trackers.getObjects().size(); j++)
            {
                rectangle(frame, trackers.getObjects()[j], Scalar(255, 0, 0), 2, 1);
            }
            imshow("tracker", frame);
        }
        if (waitKey(1) == 27)break;
    }
    return 0;
}

cv::Point 支持直接乘法

cv::Point p(100,200);

p = p*0.3; // 结果p=(30,60)

cv::Point2f ptf(0.3f, 0.7f);
cv::Point pti;
pti = ptf; # 内部采用四舍五入取整
std::cout << "cv::Point2f: " << ptf.x << ", " << ptf.y << std::endl;
std::cout << "cv::Point: " << pti.x << ", " << pti.y << std::endl;

cv::contourArea(contour) // contour.size必须>0，否则有assert报错

cv::selectROI() // 选择rect

cv::selectROIs() // 选择rect

以下代码测试了OpenCV的DFT，注意：

按照实部->[0]位置，虚部->[1]位置设置能正常运行，反之也能，为了不至于混淆，一律实部->[0]，虚部->[0]。
此例程只测试了一行的效果，实际上单独一列算出来结果也一样正确。

    const float fb = 5.1f; // 信号频率
    const float T  = 1.0f; // 总时间
    const float fs = 32.0f; // 采样率
    const float Ts = 1.0/fs; // 采样间隔
    const int real_id = 0, imag_id = 1;
    cv::Mat x = cv::Mat::zeros(1, (int)std::round(T*fs), CV_32FC2);
    cv::Mat F = x.clone();

    for (int k = 0; k < x.cols; k++)
    {
        const float t = k*Ts;
        x.at<cv::Vec2f>(0,k)[real_id] = std::cos(2*CV_PI*fb*t);
    }
    cv::dft(x, F, cv::DFT_SCALE);
    for (int k = 0; k < x.cols; k++)
    {
        printf("x(%03d) = %+08.3f + %+08.3f*i
",
               k, x.at<cv::Vec2f>(0,k)[real_id], x.at<cv::Vec2f>(0,k)[imag_id]);
    }
    for (int i = 0; i < F.cols; i++)
    {
        const int k = (i + (F.cols/2)) % F.cols;
        const float f = ((float)k/T < fs/2) ? ((float)k/T) : ((float)k/T - fs);

        const float real_part = F.at<cv::Vec2f>(0,k)[real_id];
        const float imag_part = F.at<cv::Vec2f>(0,k)[imag_id];

        printf("%03d#   F(%+08.3f) = %+08.3f + %+08.3f*i =  %+08.3f @%+08.3f度
",
               k, f, real_part, imag_part,
               cv::norm(F.at<cv::Vec2f>(0,k)), CV_PI/180.0*atan2(imag_part, real_part));
    }

读取参数文件最简洁案例如下，由于要加载、释放两次文件，所以速度比加载一次慢。当然加载一次的方案实现起来没有这么简洁。

cv::FileStorage(“file_name.yml", cv::FileStorage::READ)["parameter_1"] >> parameter_1;

cv::FileStorage(“file_name.yml", cv::FileStorage::READ)["parameter_2"] >> parameter_2;

用opencv cv::FileStorage读取xml文件时，要求xml文件根目录必须是<opencv_storage>，否则会报错

cv::fillConvexPoly(image_to_fill, one_contour, cv::Scalar(255), cv::LINE_AA); // 填充one_contour包含区域，结果在image_to_fill中

cv::resize(img_src, img_linear, cv::Size(13, 13), 0, 0, cv::INTER_LINEAR); // 双线性
cv::resize(img_src, img_area, cv::Size(13, 13), 0, 0, cv::INTER_AREA ); // 取区域平均值

OpenCV在处理+-*/时，会自动处理饱和，如下代码中，如果dat为CV_8UC1，相减后的像素如果<0，则会置为0，不用担心溢出。

    cv::Mat dat;
    // .... ohter code
    dat = dat - cv::mean(dat);