前景概要
很多机器视觉的定位与识别场景，如无人车、无人机，都会用Aruco码或特定的标志物来实现，Aruco码的优点在于，xxxx(自行搜索)。
对于像在低成本轻量级的无人机这种嵌入式系统上，搭载深度学习的识别算法目前还有困难
所以大家现在采用较多的，就是识别特定的标志物。之前看小鹏汽车的宣传片，他们也是在用黑白同心圆环的目标板。
实现内容
0、打开摄像头或某张图片
1、先检测圆环(因视角变换可能是椭圆环)；
2、裁剪保留圆环区域
3、检测圆环中的Aruco码(单个或菱形或棋盘)
4、计算目标的位姿
大致效果
代码预览
完整代码请看github（测试性代码，写的很粗糙，仅供参考）
Github: https://github.com/1061700625/OpenCV_Aruco
代码语言：javascript复制Mat testDetect(Mat &markerImage, bool diamond = true, bool aamed=false, bool show=false) {
/*************************************检测椭圆*****************************************************/
// FeaturePoint(img ,img2);
// EDCircle(markerImage);
float t1, t2, tdelt;
vector<Mat> results;
Mat fullSplitImage;
if(aamed) {
t1 = cv::getTickCount();
AAMED_Fled(markerImage, fullSplitImage, results);
t2 = cv::getTickCount();
tdelt = 1000.0*(t2-t1) / cv::getTickFrequency();
std::cout << "AAMED耗时(ms)：" << tdelt << std::endl;
}else {
results.emplace_back(markerImage);
fullSplitImage = markerImage.clone();
}
/************************************************************************************************/
// 对每个椭圆区域进行检测
for (auto& cropSplitImage: results) {
// 检测Aruco
vector<vector<Point2f>> diamondCorners;
vector<cv::Vec4i> diamondIds;
vector<vector<Point2f>> markerCorners;
vector<int> markerIds;
vector<vector<Point2f>> rejectedCandidates;
t1 = cv::getTickCount();
// cv::copyMakeBorder(cropSplitImage, cropSplitImage, 5, 5, 5, 5, cv::BORDER_CONSTANT, Scalar(255,0,0));
detectAruco(cropSplitImage, markerCorners, rejectedCandidates, markerIds);
if (markerCorners.empty()) {
cout<<"无可用Marker"<<endl;
return fullSplitImage;
}
// 显示检测到的但是由于字典对不上被拒绝的Marker
if(show) {
if (!rejectedCandidates.empty()){
cout<<"一共有 "<<rejectedCandidates.size()<<" 个被拒绝的 Marker "<<endl;
for (auto & rejectedCandidate : rejectedCandidates) {
for (int i=0;i<4;i ) {
cv::circle(fullSplitImage,cv::Point(rejectedCandidate[i].x,rejectedCandidate[i].y),6,cv::Scalar(0,0,255));
}
}
}
}
if(diamond) {
detectDiamon(cropSplitImage, markerCorners, markerIds, diamondCorners, diamondIds);
}
t2 = cv::getTickCount();
tdelt = 1000.0*(t2-t1) / cv::getTickFrequency();
std::cout << "检测耗时(ms)：" << tdelt << std::endl;
std::vector<cv::Vec3d> rvecs, tvecs;
cv::Vec3d rvec, tvec;
t1 = cv::getTickCount();
if(diamond) {
if (diamondIds.empty()) {
cout<<"无可用diamondIds"<<endl;
continue;
}
// 绘制检测边框
if(show) {
cv::aruco::drawDetectedDiamonds(fullSplitImage, diamondCorners, diamondIds);
}
// 估计相机位姿(相对于每一个marker) markerLength为什么是squareLength?
cv::aruco::estimatePoseSingleMarkers(diamondCorners, markerLength, cameraMatrix, distCoeffs, rvecs, tvecs);
}else {
if (markerIds.empty()){
cout<<"无可用markerIds"<<endl;
continue;
}
// 绘制检测边框
if(show) {
cv::aruco::drawDetectedMarkers(fullSplitImage, markerCorners, markerIds);
}
// 估计相机位姿(相对于每一个marker)
// cv::aruco::estimatePoseSingleMarkers(markerCorners, squareLength, cameraMatrix, distCoeffs, rvecs, tvecs);
// 估计相机位姿(相对于 aruco 板)
cv::aruco::estimatePoseBoard(markerCorners, markerIds, board, cameraMatrix, distCoeffs, rvec, tvec); rvecs.emplace_back(rvec); tvecs.emplace_back(tvec);
}
t2 = cv::getTickCount();
tdelt = 1000.0*(t2-t1) / cv::getTickFrequency();
std::cout << "相机位姿估计耗时(ms)：" << tdelt << std::endl;
// 为每个标记画轴
t1 = cv::getTickCount();
for (int i = 0; i < rvecs.size(); i) {
rvec = rvecs[i];
tvec = tvecs[i];
// 得到的位姿估计是：从board坐标系到相机坐标系的
cv::Mat R;
cv::Rodrigues(rvec,R);
Eigen::Matrix3d R_eigen;
cv::cv2eigen(R,R_eigen);
// Eigen中使用右乘的顺序, 因此ZYX对应的是012, 实际上这个编号跟乘法的顺序一致就可以了(从左向右看的顺序)
Eigen::Vector3d zyx_Euler_fromR = R_eigen.eulerAngles(0,1,2);
if(show) {
cout << "R_{camera<---marker} :" << R << endl;
cout << "t_{camera<---marker} :" << tvec << endl;
cout << "zyx旋转欧拉角[输出顺序为:x,y,z]: " << (180)/(M_PI)*zyx_Euler_fromR.transpose()<<endl;
cv::aruco::drawAxis(fullSplitImage, cameraMatrix, distCoeffs, rvec, tvec, 0.1);
cout << "--------------------------------------------" << endl;
}
//cv::aruco::drawAxis(markerImage, cameraMatrix, distCoeffs, rvec, tvec, 0.1);
}
t2 = cv::getTickCount();
tdelt = 1000.0*(t2-t1) / cv::getTickFrequency();
std::cout << "欧拉角耗时(ms)：" << tdelt << std::endl;
}
if (aamed && show) {
imshow("markerImage", markerImage);
//imshow("fullSplitImage", fullSplitImage);
//imwrite("../fullSplitImage.png", fullSplitImage);
}
// waitKey(0);
return fullSplitImage;
}
int main(int argc, char* argv[])
{
Mat fullSplitImage;
Mat boardImage;
board->draw( cv::Size(200, 200), // 整个board的大小
boardImage, // 返回的图像
10, // 整个board的边距
1 ); // 每个码内的边距
imwrite("../boardImage.png", boardImage);
cv::aruco::drawCharucoDiamond(dictionary, cv::Vec4i(0,1,2,3), 200, 150, boardImage);
imwrite("../diamondImage.png", boardImage);
//vector<Mat> markerImages = generateAruco(5);
//Mat markerImage = markerImages[0];
cout<<">> 预处理完成!"<<endl;
/************************************************************************************************/
float t1, t2, tdelt;
Mat markerImage = imread("../img/4.jpg", 1);
t1 = cv::getTickCount();
fullSplitImage = testDetect(markerImage, false, false, true);
t2 = cv::getTickCount();
tdelt = 1000.0*(t2-t1) / cv::getTickFrequency();
std::cout << "一帧总耗时(ms)：" << tdelt << std::endl;
imshow("fullSplitImage", fullSplitImage);
waitKey(0);
return 0;
/************************************************************************************************/
Mat frame;
VideoCapture capture;
capture.open(0);
if (!capture.isOpened()) {
cerr << "ERROR! Unable to open cameran";
return -1;
}
cv::namedWindow("fullSplitImage",0);
cv::resizeWindow("fullSplitImage", 960, 540);
for (;;) {
capture.read(frame);
cv::resize(frame, frame, cv::Size(752, 480), 0, 0, cv::INTER_AREA);
cout << frame.size << endl;
if (frame.empty()) {
cerr << "ERROR! blank frame grabbedn";
break;
}
t1 = cv::getTickCount();
fullSplitImage = testDetect(frame, false, false);
t2 = cv::getTickCount();
tdelt = 1000.0*(t2-t1) / cv::getTickFrequency();
std::cout << "一帧耗时(ms)：" << tdelt << std::endl;
imshow("fullSplitImage", fullSplitImage.empty() ? frame : fullSplitImage);
if (waitKey(1) >= 0)
break;
}
return 0;
}其他内容
- 在线aruco标记生成器：https://chev.me/arucogen/
- OpenCV识别Aruco markers库：https://docs.opencv.org/4.5.4/d5/dae/tutorial_aruco_detection.html
