配置Json环境

使用Jsoncpp包中的.cpp和 .h文件
解压上面下载的 Jsoncpp 文件，把 jsoncpp-src-0.5.0文件拷贝到工程目录下，将 jsoncpp-src-0.5.0\jsoncpp-src-0.5.0\include\json 和 jsoncpp-src-0.5.0\jsoncpp-src-0.5.0\src\lib_json 目录里的文件包含到VS工程中，在VS工程的属性C/C++下General中 Additional Include Directories 包含头文件目录 .\jsoncpp-src-0.5.0\include 。在使用的cpp文件中包含json头文件即可，如： #include "json/json.h" 。将 json_reader.cpp、 json_value.cpp和 json_writer.cpp三个文件的Precompiled Header属性设置为 Not Using Precompiled Headers，否则编译会出现错误。

试了很多方法，上述方法亲测可用。

准备工作

读图解析Json文件及声明变量

    Mat src = imread(".//images//chessboard.jpg");//声明一些存储直线点的变量int i=0;vector < vector<Point2d> > Points;vector<Point2d> XZ_line, YZ_line, XY_line, XZ_line1, YZ_line1, XY_line1;Points.push_back(XZ_line);Points.push_back(YZ_line);Points.push_back(XY_line);Points.push_back(XZ_line1);Points.push_back(YZ_line1);Points.push_back(XY_line1);Point2d orignPoint;//第一步先解析Json数据//用来判断是线还是交点的点string shape_type = "line";//用来区分线条的类型vector<string> linetype;string c = ",";linetype.push_back("XZ_line");linetype.push_back("YZ_line");linetype.push_back("XY_line");linetype.push_back("XZ_line'");linetype.push_back("YZ_line'");linetype.push_back("XY_line'");//声明读取Json文件变量Json::Reader reader;Json::Value root;ifstream in(".//labelme_data//chessboard_line.json", ios::binary);if (!in.is_open()){cout << "打开文件错误\n";return 0;}if (reader.parse(in,root)){int temp = root["shapes"].size();for (int m =0; m < temp; m++){string shapetype = root["shapes"][m]["shape_type"].asString();//判断是否为直线点，还是交点if (shapetype == shape_type){string templinetype = root["shapes"][m]["label"].asString();for (; i < linetype.size(); i++){if (templinetype == linetype[i]){//得到points标签下的数据（此时为string类型）int temppointsize = root["shapes"][m]["points"].size();for (int j = 0; j < temppointsize; j++){string strpoint = root["shapes"][m]["points"][j].toStyledString();//对读取到的字符串进行处理，删除[]和空格字符strpoint.erase(strpoint.find(' '),1);strpoint.erase(strpoint.find('['), 1);strpoint.erase(strpoint.find(' '), 1);strpoint.erase(strpoint.find(']'), 1);//将该字符串按逗号分割vector<string> res;SplitString(strpoint, res, c);//将字符串转为doublePoint2d temppoint;temppoint.x = atof(res[0].c_str());temppoint.y = atof(res[1].c_str());Points[i].push_back(temppoint);}}}i = 0;}//原点位置else if (shapetype == "point"){//得到points标签下的数据（此时为string类型）string strpoint = root["shapes"][m]["points"].toStyledString();//对读取到的字符串进行处理，删除[]和空格字符strpoint.erase(strpoint.find(' '), 1);strpoint.erase(strpoint.find('['), 1);strpoint.erase(strpoint.find('['), 1);strpoint.erase(strpoint.find(' '), 1);strpoint.erase(strpoint.find(']'), 1);strpoint.erase(strpoint.find(']'), 1);//将该字符串按逗号分割vector<string> res;SplitString(strpoint, res, c);//将字符串转为doubleorignPoint.x = atof(res[0].c_str());orignPoint.y = atof(res[1].c_str());}}in.close();}

读取到的图像如下：

将Json文件里面标记的直线和点加载到图片上，并绘制出来可视化

    //点读取完毕，根据点将线条画在图上//复制一张图像//YZ_line, XY_line, XZ_line1, YZ_line1, XY_line1;Mat src1 = src.clone();for (int k = 0; k < Points.size(); k++){int tempSize = Points[k].size()/2;for (int n = 0; n < tempSize; n++){switch (k){case 0:line(src1, Points[k][2 * n], Points[k][2 * n + 1], Scalar(0, 0, 255), 1);XZ_line.push_back(Points[k][2 * n]);XZ_line.push_back(Points[k][2 * n+1]);break;case 1:line(src1, Points[k][2 * n], Points[k][2 * n + 1], Scalar(0, 255, 0), 1);YZ_line.push_back(Points[k][2 * n]);YZ_line.push_back(Points[k][2 * n + 1]);break;case 2:line(src1, Points[k][2 * n], Points[k][2 * n + 1], Scalar(255, 0, 0), 1);XY_line.push_back(Points[k][2 * n]);XY_line.push_back(Points[k][2 * n + 1]);break;case 3:line(src1, Points[k][2 * n], Points[k][2 * n + 1], Scalar(0, 255, 255), 1);XZ_line1.push_back(Points[k][2 * n]);XZ_line1.push_back(Points[k][2 * n + 1]);break;case 4:line(src1, Points[k][2 * n], Points[k][2 * n + 1], Scalar(255, 255, 0), 1);YZ_line1.push_back(Points[k][2 * n]);YZ_line1.push_back(Points[k][2 * n + 1]);break;case 5:line(src1, Points[k][2 * n], Points[k][2 * n + 1], Scalar(255, 0, 255), 1);XY_line1.push_back(Points[k][2 * n]);XY_line1.push_back(Points[k][2 * n + 1]);break;default:break;}}}

加载完，不同平面的直线用不同颜色绘制的直线和点如下图所示：

计算影消点

计算影消点，就是根据上面各个平面平行的直线，拟合出平行直线的交点。平行的直线本该没有交点，但经过摄影几何的透视变换，现实世界坐标系下的平行线都会相交于一点，这个点就是影消点，也就是世界坐标系里面的无穷远点，两条平行直线相交于无穷远处。但在摄影几何下，经过透视变换，无穷远点变为影消点，影消点在图像坐标系下可以求。

无穷远点变为影消点最直观的例子，拍摄一条火车铁轨，铁轨最终在图像里面一个点相交。

    //接下来求图像上的影消点//即计算两个平面平行线的交点//先计算XZ平面，两条平行线的交点//计算出来的line为Vec4f类型，即line[0-3]//斜率为line[1]/line[0],并过点(line[2],line[3])//先将其转为向量相乘的形式即l·x=0的形式，这里l其实为l的转置//l为(line[1],-line[0],-line[1]*line[2]+line[0]*line[3])vector<Vec3d> vecLineXZ;vector<Vec3d> vecLineYZ;vector<Vec3d> vecLineXY;vector<Vec3d> vecLineXZ1;vector<Vec3d> vecLineYZ1;vector<Vec3d> vecLineXY1;for (int i = 0; i < vecXZLine.size(); i++){Vec3d tempvec;tempvec[0] = vecXZLine[i][1];tempvec[1] = -vecXZLine[i][0];tempvec[2] = -vecXZLine[i][1] * vecXZLine[i][2] + vecXZLine[i][0] * vecXZLine[i][3];vecLineXZ.push_back(tempvec);}for (int i = 0; i < vecYZLine.size(); i++){Vec3d tempvec;tempvec[0] = vecYZLine[i][1];tempvec[1] = -vecYZLine[i][0];tempvec[2] = -vecYZLine[i][1] * vecYZLine[i][2] + vecYZLine[i][0] * vecYZLine[i][3];vecLineYZ.push_back(tempvec);}for (int i = 0; i < vecXYLine.size(); i++){Vec3d tempvec;tempvec[0] = vecXYLine[i][1];tempvec[1] = -vecXYLine[i][0];tempvec[2] = -vecXYLine[i][1] * vecXYLine[i][2] + vecXYLine[i][0] * vecXYLine[i][3];vecLineXY.push_back(tempvec);}for (int i = 0; i < vecXZ1Line.size(); i++){Vec3d tempvec;tempvec[0] = vecXZ1Line[i][1];tempvec[1] = -vecXZ1Line[i][0];tempvec[2] = -vecXZ1Line[i][1] * vecXZ1Line[i][2] + vecXZ1Line[i][0] * vecXZ1Line[i][3];vecLineXZ1.push_back(tempvec);}for (int i = 0; i < vecYZ1Line.size(); i++){Vec3d tempvec;tempvec[0] = vecYZ1Line[i][1];tempvec[1] = -vecYZ1Line[i][0];tempvec[2] = -vecYZ1Line[i][1] * vecYZ1Line[i][2] + vecYZ1Line[i][0] * vecYZ1Line[i][3];vecLineYZ1.push_back(tempvec);}for (int i = 0; i < vecXY1Line.size(); i++){Vec3d tempvec;tempvec[0] = vecXY1Line[i][1];tempvec[1] = -vecXY1Line[i][0];tempvec[2] = -vecXY1Line[i][1] * vecXY1Line[i][2] + vecXY1Line[i][0] * vecXY1Line[i][3];vecLineXY1.push_back(tempvec);}//求XZ平面的影消点Vec3d testVanishPoint = RobertColins(XZ_line, src.cols, src.rows);Vec3d pointvanishXZvec = vecLineXZ[0].cross(vecLineXZ[vecLineXZ.size() - 1]);//将齐次点形式转为欧式坐标Point2d pointvanishXZ = Point2d(pointvanishXZvec[0] / pointvanishXZvec[2], pointvanishXZvec[1] / pointvanishXZvec[2]);//将图片放大，将影消点可视化Mat dst(6060, 6060, CV_8UC3, Scalar(255, 255, 255));Rect roiRect = Rect(2827, 2827, src.cols, src.rows);src1.copyTo(dst(roiRect));//定义偏移量Point2d dispoint = Point2d(2828, 2828);line(dst, XZ_line[0] + dispoint, pointvanishXZ + dispoint, Scalar(0, 0, 255), 1);line(dst, XZ_line[XZ_line.size() - 1] + dispoint, pointvanishXZ + dispoint, Scalar(0, 0, 255), 1);//求YZ平面的影消点Vec3d pointvanishYZvec = vecLineYZ[0].cross(vecLineYZ[vecLineYZ.size() - 1]);Point2d pointvanishYZ = Point2d(pointvanishYZvec[0] / pointvanishYZvec[2], pointvanishYZvec[1] / pointvanishYZvec[2]);line(dst, YZ_line[1] + dispoint, pointvanishYZ + dispoint, Scalar(0, 255, 0), 1);line(dst, YZ_line[YZ_line.size() - 1] + dispoint, pointvanishYZ + dispoint, Scalar(0, 255, 0), 1);//求XY平面的影消点Vec3d pointvanishXYvec = vecLineXY[0].cross(vecLineXY[vecLineXY.size() - 1]);Point2d pointvanishXY = Point2d(pointvanishXYvec[0] / pointvanishXYvec[2], pointvanishXYvec[1] / pointvanishXYvec[2]);line(dst, XY_line[1] + dispoint, pointvanishXY + dispoint, Scalar(255, 0, 0), 1);line(dst, XY_line[XY_line.size() - 1] + dispoint, pointvanishXY + dispoint, Scalar(255, 0, 0), 1);//求XZ'平面的影消点Vec3d pointvanishXZ1vec = vecLineXZ1[0].cross(vecLineXZ1[vecLineXZ1.size() - 1]);Point2d pointvanishXZ1 = Point2d(pointvanishXZ1vec[0] / pointvanishXZ1vec[2], pointvanishXZ1vec[1] / pointvanishXZ1vec[2]);line(dst, XZ_line1[0] + dispoint, pointvanishXZ1 + dispoint, Scalar(0, 255, 255), 1);line(dst, XZ_line1[XZ_line1.size() - 1] + dispoint, pointvanishXZ1 + dispoint, Scalar(0, 255, 255), 1);//求YZ'平面的影消点Vec3d pointvanishYZ1vec = vecLineYZ1[0].cross(vecLineYZ1[vecLineYZ1.size() - 1]);Point2d pointvanishYZ1 = Point2d(pointvanishYZ1vec[0] / pointvanishYZ1vec[2], pointvanishYZ1vec[1] / pointvanishYZ1vec[2]);line(dst, YZ_line1[1] + dispoint, pointvanishYZ1 + dispoint, Scalar(255, 255, 0), 1);line(dst, YZ_line1[YZ_line1.size() - 1] + dispoint, pointvanishYZ1 + dispoint, Scalar(255, 255, 0), 1);//求XY'平面的影消点Vec3d pointvanishXY1vec = vecLineXY1[0].cross(vecLineXY1[vecLineXY1.size() - 1]);Point2d pointvanishXY1 = Point2d(pointvanishXY1vec[0] / pointvanishXY1vec[2], pointvanishXY1vec[1] / pointvanishXY1vec[2]);line(dst, XY_line1[1] + dispoint, pointvanishXY1 + dispoint, Scalar(255, 0, 255), 1);line(dst, XY_line1[XY_line1.size() - 1] + dispoint, pointvanishXY1 + dispoint, Scalar(255, 0, 255), 1);//至此，各个平面的影消点都已求出来//根据XZ、YZ、XY平面的影消点，进行相机内参标定，求解W矩阵vector<Point2d> vanishPointVec;vanishPointVec.push_back(pointvanishXZ);vanishPointVec.push_back(pointvanishYZ);vanishPointVec.push_back(pointvanishXY);return 0;
}

求得影消点的图片如下图所示：

计算影消点时，用到了柯林斯算法，该算法如下：

//罗伯特·柯林斯方法求影消点
//输入参数为直线的端点，图像的宽高
//输出为影消点的齐次坐标形式
Vec3d RobertColins(vector<Point2d> vecLinePoints,int cols,int rows)
{Vec3d vanishPoint;int size = vecLinePoints.size();//防止坐标点差异过大造成求解精度下降，进行偏移int offset = (cols + rows) / 4;//这里直线端点，以两个点为一组代表一条直线//因此size必为偶数//得出各条直线的方向向量vector <Vec3d> Lines;for (int i = 0; i < size/2; i++){Vec3d startPoint, endPoint;startPoint[0] = (vecLinePoints[2 * i].x-offset);startPoint[1] = (vecLinePoints[2 * i].y-offset);startPoint[2] = offset;endPoint[0] = (vecLinePoints[2 * i + 1].x - offset);endPoint[1] = (vecLinePoints[2 * i + 1].y - offset);endPoint[2] = offset;//两个点坐标向量的叉乘等于，过该两点的直线齐次坐标向量Lines.push_back(startPoint.cross(endPoint));}/*声明一个Mat M形成3×3的“二阶矩”矩阵M为[a_i*a_i        a_i*b_i        a_i*c_i]M = sum [a_i*b_i        b_i*b_i        c_i*b_i][a_i*c_i        b_i*c_i        c_i*c_i]其中取i = 1到n的和。注意，M是对称矩阵。*/Mat M = (Mat_<double>(3, 3)<<0,0,0,0,0,0,0,0,0);for (int i = 0; i < size / 2; i++){Mat tempM = (Mat_<double>(3, 3));tempM.at<double>(0, 0) = Lines[i][0] * Lines[i][0];tempM.at<double>(0, 1) = Lines[i][0] * Lines[i][1];tempM.at<double>(0, 2) = Lines[i][0] * Lines[i][2];tempM.at<double>(1, 0) = Lines[i][1] * Lines[i][0];tempM.at<double>(1, 1) = Lines[i][1] * Lines[i][1];tempM.at<double>(1, 2) = Lines[i][1] * Lines[i][2];tempM.at<double>(2, 0) = Lines[i][2] * Lines[i][0];tempM.at<double>(2, 1) = Lines[i][2] * Lines[i][1];tempM.at<double>(2, 2) = Lines[i][2] * Lines[i][2];M = M + tempM;}//对M矩阵进行特征值分解Mat eigenvector = (Mat_<double>(3, 3));Mat eigenvalue = (Mat_<double>(3, 1));eigen(M, eigenvalue, eigenvector);vanishPoint[0] = eigenvector.at<double>(0, 2)*(offset / (eigenvector.at<double>(0, 2))) + offset;vanishPoint[1] = eigenvector.at<double>(1, 2)*(offset / (eigenvector.at<double>(0, 2))) + offset;vanishPoint[2] = 1;return vanishPoint;
}

解析Json文件用到的字符串处理方法：

//分割points字符串类型的点
void SplitString(const string& s, vector<string>& v, const string& c)
{string::size_type pos1, pos2;pos2 = s.find(c);pos1 = 0;while (string::npos != pos2){v.push_back(s.substr(pos1, pos2 - pos1));pos1 = pos2 + c.size();pos2 = s.find(c, pos1);}if (pos1 != s.length())v.push_back(s.substr(pos1));
}

鲁鹏老师三维重建课程之单视图重建相关推荐

北邮鲁鹏老师三维重建课程之相机标定
在看北邮鲁鹏老师的三维重建的课程过程中,去官网找到有三个作业.现将三个作业里面的第一个作业相机标定完成.总体来说,可以分为三个部分,即图像坐标点和世界坐标点的获取:映射矩阵的生成,相机内外参的求解三个 ...
【计算机视觉】计算机视觉与深度学习-北邮鲁鹏老师课程笔记
计算机视觉与深度学习-北邮鲁鹏老师课程笔记 01-计算机视觉相关介绍 02-图像分类 03-全连接神经网络 04-图像去噪&卷积&边缘提取 05-纹理表示&卷积神经网络 06- ...
【计算机视觉】计算机视觉与深度学习-01-计算机视觉相关介绍-北邮鲁鹏老师课程笔记
计算机视觉与深度学习-01-计算机视觉相关介绍前言图像处理 vs 计算机视觉图像处理计算机视觉相关课程计算机视觉简介计算机视觉顶级会议计算机视觉的目标图像中的信息三维场景的结构信息 ...
北邮鲁鹏老师视频笔记
https://www.zhihu.com/people/xu-hao-68-3/posts https://blog.csdn.net/weixin_45562510/category_109604 ...
影消点、影消线与相机内参、平面法向量的推导—单视图重构
影消点计算算法如下: 单视图几何Vanish Point(消失点/灭点)计算方法--Robert_T_Collins(罗伯特·柯林斯)算法https://blog.csdn.net/beyond951 ...
鲁鹏计算机视觉与深度学习
本笔记来自于鲁鹏老师的视频 https://www.bilibili.com/video/BV1V54y1B7K3?p=3&vd_source=4ddee6259222ad3b06bedffc ...
DeepFusion：基于单视图深度和梯度预测的单目SLAM实时稠密三维重建
点击上方"3D视觉工坊",选择"星标" 干货第一时间送达标题:DeepFusion: Real-Time Dense 3D Reconstruction fo ...
单视图测量 (2D变换、影消点线、单视图重构)
写在前面:本篇Blog仅作为学习笔记,学习内容来自于北邮CV-XUEBA团队的三维重建(精简版,鲁鹏)课程. 回顾经典2D变换等距变换旋转矩阵(Rotate Matrix)的性质分析证明:旋转矩 ...
鲁祥老师吉他课学习笔记
鲁祥老师吉他课学习笔记导语参考教材: <吉他入门经典教程>李国标 <弹指之间>潘尚文 <吉他教本>好连得出版社 <吉他教程>杰瑞吉他学校 <伯 ...

鲁鹏老师三维重建课程之单视图重建

配置Json环境

准备工作

读图解析Json文件及声明变量

计算影消点

鲁鹏老师三维重建课程之单视图重建相关推荐

最新文章

热门文章