注意到位姿節點之間的變換並不是位姿,之前一直有誤解;
一般地;
路標節點:也就是觀測方程【數學形式下見】的觀測值,也就是特征點的像素坐標[u,v],或者該幀相機坐標系下的3d坐標[x,y,z];
位姿節點:也就是運動方程【數學形式下見】的輸出值。例如:上述x1、x2、x3、X4對應位姿為:Tcw1、Tcw2、Tcw3、Tcw4。這里的Tcw表示對應幀相機坐標系->世界坐標系的變換;比如:在x1處看到了路標點p1,在x2處也看到了路標點p1(當然是通過特征匹配才知道再次看到)考慮以下兩種情況:
1> x1對應第一幀,x2對應第二幀
那么x1的相機坐標系即為:世界坐標系;那么第一幀位姿我們直接初始化為:
第二幀我們要:通過……solvePNPransac(points3d_1, points2d_2,R,t);來求解,points3d_1對應的是世界坐標系下的3D點(在這種情形,第一幀相機坐標系也就是世界坐標系,所以第一幀相機坐標系下的3D點), points2d_2是第二幀像素點(設第二幀對應3D路標點(在第二幀相機坐標系下)為:points3d_2)。
在這里[左乘],
Tcw2 * points3d_1 = points3d_2
同時必須知道:(假設兩幀之間的變換 T12,也就是 位姿變換)
T12 * Tcw1 = Tcw2
2> x2對應第2幀,x3對應第3幀
接着,我們再次想通過solvePNPransac(points3d_2_, points2d_3,R,t)來求解Tcw3,points3d_2_ (世界坐標系)肯定不是 points3d_2(相機坐標系),所以在第一種情形你就應該提前將points3d_2 從它所在的相機坐標系映射到世界坐標系,故:
points3d_2_ = Tcw2.inverse() * points3d_2
我們得出流程圖如下圖:
觀測模型:具體見下
運動模型:具體見下
兩個模型與G2O結合了解說明;
位姿:位置和姿態;
對於上述位姿圖,我們假設 x1 對應參考幀 ref_(同時假設是第一幀),x2對應 當前幀curr_(同時假設是第二幀);我們通過......最后通過opencv中API solvePNPransac();求出的R|t 其實是Tcw;
Tcw指的是:當前幀與世界坐標系之間的變換;
Trw指的是:參考幀與世界坐標系之間的變換。那么待估計的兩幀之間的變換為:Tcr,構成左乘關系:
Tcr, s.t. Tcr * Trw = Tcw
P_camera_curr 指的是:當前幀3D路標點 (相機坐標系)
P_world_ref 指的是:參考幀3D路標點 (世界坐標系)
P_camera_curr = Tcw*P_world_ref
這里 opencv中的 solvePNPransac 求出來的是 T = (R|t);依據上述公式:所以你要將當前幀3D路標點轉化到世界坐標系,應該是左乘:Tcw.inverse() * P_camera_curr = P_world_ref;請看以下代碼【就是匹配,第一幀相機坐標系當作世界坐標系,第二幀就是相機坐標系,重點關注代碼 81行 - 124行】:

1 #include <iostream> 2 #include <opencv2/core/core.hpp> 3 #include <opencv2/features2d/features2d.hpp> 4 #include <opencv2/highgui/highgui.hpp> 5 #include <opencv2/calib3d/calib3d.hpp> 6 #include <Eigen/Core> 7 #include <Eigen/Geometry> 8 #include <g2o/core/base_vertex.h> 9 #include <g2o/core/base_unary_edge.h> 10 #include <g2o/core/block_solver.h> 11 #include <g2o/core/optimization_algorithm_levenberg.h> 12 #include <g2o/solvers/csparse/linear_solver_csparse.h> 13 #include <g2o/types/sba/types_six_dof_expmap.h> 14 #include <chrono> 15 16 using namespace std; 17 using namespace cv; 18 19 #include"sophus\se3.h" 20 #include"sophus\so3.h" 21 22 void find_feature_matches( 23 const Mat& img_1, const Mat& img_2, 24 std::vector<KeyPoint>& keypoints_1, 25 std::vector<KeyPoint>& keypoints_2, 26 std::vector< DMatch >& matches); 27 28 // 像素坐標轉相機歸一化坐標 29 Point2d pixel2cam(const Point2d& p, const Mat& K); 30 31 void bundleAdjustment( 32 const vector<Point3f> points_3d, 33 const vector<Point2f> points_2d, 34 const Mat& K, 35 Mat& R, Mat& t 36 ); 37 38 int main() 39 { 40 //-- 讀取圖像 41 Mat img_1 = imread("color/b.png", CV_LOAD_IMAGE_COLOR); 42 Mat img_2 = imread("color/c.png", CV_LOAD_IMAGE_COLOR); 43 44 vector<KeyPoint> keypoints_1, keypoints_2; 45 vector<DMatch> matches; 46 find_feature_matches(img_1, img_2, keypoints_1, keypoints_2, matches); 47 cout << "一共找到了" << matches.size() << "組匹配點" << endl; 48 49 // 建立3D點 50 Mat depth_1 = imread("depth/b.pgm", CV_LOAD_IMAGE_UNCHANGED); // 深度圖為16位無符號數,單通道圖像 51 Mat depth_2 = imread("depth/c.pgm", CV_LOAD_IMAGE_UNCHANGED); // 深度圖為16位無符號數,單通道圖像 52 //Mat K = (Mat_<double>(3, 3) << 520.9, 0, 325.1, 0, 521.0, 249.7, 0, 0, 1); 53 Mat K = (Mat_<double>(3, 3) << 518.0, 0, 325.5, 0, 519.0, 253.5, 0, 0, 1); 54 vector<Point3f> pts_3d1;//世界坐標點 55 vector<Point3f> pts_3d2; 56 vector<Point2f> pts_2d;//相機坐標系點(歸一化???) 57 //【注:】這里表示同一個3D點在兩幅圖中的圖二對應像素點,本來就是求世界坐標系下的r、t 58 for (DMatch m : matches) 59 { 60 //取出深度值,此時彩圖和深度圖已經對齊 61 ushort d1 = depth_1.ptr<unsigned short>(int(keypoints_1[m.queryIdx].pt.y))[int(keypoints_1[m.queryIdx].pt.x)]; 62 ushort d2 = depth_2.ptr<unsigned short>(int(keypoints_2[m.trainIdx].pt.y))[int(keypoints_2[m.trainIdx].pt.x)]; 63 if ((d1 == 0) || (d2 == 0)) // bad depth 64 continue; 65 float dd1 = d1 / 1.0; 66 float dd2 = d2 / 1.0; 67 Point2d p1 = pixel2cam(keypoints_1[m.queryIdx].pt, K); // 相機坐標系歸一化平面:p1的形式 [x/z,y/z,1] 68 Point2d p2 = pixel2cam(keypoints_2[m.trainIdx].pt, K); 69 pts_3d1.push_back(Point3f(p1.x*dd1, p1.y*dd1, dd1));//乘以Z就得到形式 [x,y,z]?????左邊相機為參考,把他當作世界坐標系,確實是這樣的 70 pts_3d2.push_back(Point3f(p2.x*dd2, p2.y*dd2, dd2)); 71 pts_2d.push_back(keypoints_2[m.trainIdx].pt);//圖二中的2D像素坐標[u,v] 72 } 73 74 cout << "3d-2d pairs: " << pts_3d1.size() << endl; 75 76 Mat r, t; 77 // 調用OpenCV 的 PnP 求解r t,可選擇EPNP,DLS等方法 78 //solvePnP(pts_3d, pts_2d, K, Mat(), r, t, false); 79 //solvePnP(pts_3d1, pts_2d, K, Mat(), r, t, false, SOLVEPNP_EPNP); 80 81 cv::solvePnPRansac(pts_3d1, pts_2d, K, Mat(), r, t, false, 100, 1.0, 0.99); 82 83 Mat R; 84 cv::Rodrigues(r, R); // r為旋轉向量形式,用Rodrigues公式轉換為矩陣 85 //RR = -R.inv(); 86 cout << "R=" << endl << R << endl; 87 cout << "t=" << endl << t << endl; 88 //************************************************************************************************ 89 // 李群 90 //************************************************************************************************ 91 Sophus::SE3 T_se3 = Sophus::SE3( 92 Sophus::SO3(r.at<double>(0, 0), r.at<double>(1, 0), r.at<double>(2, 0)), 93 Sophus::Vector3d(t.at<double>(0, 0), t.at<double>(1, 0), t.at<double>(2, 0)) 94 ); 95 cout << "T_se3 = " << T_se3.matrix() << endl; 96 //************************************************************************************************ 97 // 歐氏群 98 //************************************************************************************************ 99 Eigen::Isometry3d T_Isometry3d = Eigen::Isometry3d::Identity(); 100 //Eigen::AngleAxisd rot_vec(r.at<double>(0, 0), r.at<double>(1, 0), r.at<double>(2, 0)); 101 Eigen::Vector3d trans(t.at<double>(0, 0), t.at<double>(1, 0), t.at<double>(2, 0)); 102 Eigen::Matrix<double, 3, 3> R_; 103 R_ << R.at<double>(0,0), R.at<double>(0,1), R.at<double>(0,2), 104 R.at<double>(1,0), R.at<double>(1,1), R.at<double>(1,2), 105 R.at<double>(2,0), R.at<double>(2,1), R.at<double>(2,2); 106 T_Isometry3d.rotate(R_); 107 T_Isometry3d.pretranslate(trans); 108 cout << " T_Isometry3d = " << T_Isometry3d.matrix() << endl; 109 110 cout << "calling bundle adjustment" << endl; 111 bundleAdjustment(pts_3d1, pts_2d, K, R, t); 112 //************************************************************************************************ 113 // 計算誤差:error = point3d_curr -( R * point3d_ref + t ) 114 //************************************************************************************************ 115 for (int i = 0; i < 20; i++) 116 { 117 cout << " pts_3d1 = " << pts_3d1[i] << endl; 118 cout << " pts_3d2 = " << pts_3d2[i] << endl; 119 cout << " error " << 120 (Mat_<double>(3, 1) << pts_3d2[i].x, pts_3d2[i].y, pts_3d2[i].z) 121 - R*((Mat_<double>(3, 1) << pts_3d1[i].x, pts_3d1[i].y, pts_3d1[i].z)) - t 122 << endl; 123 cout << endl; 124 } 125 126 waitKey(0); 127 return 0; 128 } 129 130 void find_feature_matches(const Mat& img_1, const Mat& img_2, 131 std::vector<KeyPoint>& keypoints_1, 132 std::vector<KeyPoint>& keypoints_2, 133 std::vector< DMatch >& matches) 134 { 135 //-- 初始化 136 Mat descriptors_1, descriptors_2; 137 // used in OpenCV3 138 Ptr<FeatureDetector> detector = ORB::create(1000); 139 Ptr<DescriptorExtractor> descriptor = ORB::create(); 140 // use this if you are in OpenCV2 141 // Ptr<FeatureDetector> detector = FeatureDetector::create ( "ORB" ); 142 // Ptr<DescriptorExtractor> descriptor = DescriptorExtractor::create ( "ORB" ); 143 Ptr<DescriptorMatcher> matcher = DescriptorMatcher::create("BruteForce-Hamming"); 144 //-- 第一步:檢測 Oriented FAST 角點位置 145 detector->detect(img_1, keypoints_1); 146 detector->detect(img_2, keypoints_2); 147 148 //-- 第二步:根據角點位置計算 BRIEF 描述子 149 descriptor->compute(img_1, keypoints_1, descriptors_1); 150 descriptor->compute(img_2, keypoints_2, descriptors_2); 151 152 //-- 第三步:對兩幅圖像中的BRIEF描述子進行匹配,使用 Hamming 距離 153 vector<DMatch> match; 154 // BFMatcher matcher ( NORM_HAMMING ); 155 matcher->match(descriptors_1, descriptors_2, match); 156 157 //-- 第四步:匹配點對篩選 158 double min_dist = 10000, max_dist = 0; 159 160 //找出所有匹配之間的最小距離和最大距離, 即是最相似的和最不相似的兩組點之間的距離 161 for (int i = 0; i < descriptors_1.rows; i++) 162 { 163 double dist = match[i].distance; 164 if (dist < min_dist) min_dist = dist; 165 if (dist > max_dist) max_dist = dist; 166 } 167 168 printf("-- Max dist : %f \n", max_dist); 169 printf("-- Min dist : %f \n", min_dist); 170 171 //當描述子之間的距離大於兩倍的最小距離時,即認為匹配有誤.但有時候最小距離會非常小,設置一個經驗值30作為下限. 172 for (int i = 0; i < descriptors_1.rows; i++) 173 { 174 if (match[i].distance <= max(2 * min_dist, 30.0)) 175 { 176 matches.push_back(match[i]); 177 } 178 } 179 180 181 182 Mat img_match; 183 Mat img_goodmatch; 184 drawMatches(img_1, keypoints_1, img_2, keypoints_2, match, img_match); 185 drawMatches(img_1, keypoints_1, img_2, keypoints_2, matches, img_goodmatch); 186 namedWindow("所有匹配點對", CV_WINDOW_NORMAL); 187 imshow("所有匹配點對", img_match); 188 namedWindow("優化后匹配點對", CV_WINDOW_NORMAL); 189 imshow("優化后匹配點對", img_goodmatch); 190 } 191 192 Point2d pixel2cam(const Point2d& p, const Mat& K)// [u,v,1] - > [x/z, y/z, 1] 193 { 194 return Point2d 195 ( 196 (p.x - K.at<double>(0, 2)) / K.at<double>(0, 0), 197 (p.y - K.at<double>(1, 2)) / K.at<double>(1, 1) 198 ); 199 } 200 201 202 void bundleAdjustment( 203 const vector< Point3f > points_3d, 204 const vector< Point2f > points_2d, 205 const Mat& K, 206 Mat& R, Mat& t) 207 { 208 // 初始化g2o 209 typedef g2o::BlockSolver< g2o::BlockSolverTraits<6, 3> > Block; // pose 維度為 6, landmark 維度為 3 210 Block::LinearSolverType* linearSolver = new g2o::LinearSolverCSparse<Block::PoseMatrixType>(); // 線性方程求解器 211 Block* solver_ptr = new Block(linearSolver); // 矩陣塊求解器 212 g2o::OptimizationAlgorithmLevenberg* solver = new g2o::OptimizationAlgorithmLevenberg(solver_ptr); 213 g2o::SparseOptimizer optimizer; 214 optimizer.setAlgorithm(solver); 215 216 // vertex 217 g2o::VertexSE3Expmap* pose = new g2o::VertexSE3Expmap(); // camera pose 218 Eigen::Matrix3d R_mat; 219 R_mat << 220 R.at<double>(0, 0), R.at<double>(0, 1), R.at<double>(0, 2), 221 R.at<double>(1, 0), R.at<double>(1, 1), R.at<double>(1, 2), 222 R.at<double>(2, 0), R.at<double>(2, 1), R.at<double>(2, 2); 223 pose->setId(0); 224 pose->setEstimate(g2o::SE3Quat( 225 R_mat, 226 Eigen::Vector3d(t.at<double>(0, 0), t.at<double>(1, 0), t.at<double>(2, 0)) 227 )); 228 optimizer.addVertex(pose); 229 230 int index = 1; 231 for (const Point3f p : points_3d) // landmarks 232 { 233 g2o::VertexSBAPointXYZ* point = new g2o::VertexSBAPointXYZ(); 234 point->setId(index++); 235 point->setEstimate(Eigen::Vector3d(p.x, p.y, p.z)); 236 point->setMarginalized(true); // g2o 中必須設置 marg 參見第十講內容 237 optimizer.addVertex(point); 238 } 239 240 // parameter: camera intrinsics 241 g2o::CameraParameters* camera = new g2o::CameraParameters( 242 K.at<double>(0, 0), Eigen::Vector2d(K.at<double>(0, 2), K.at<double>(1, 2)), 0 243 ); 244 camera->setId(0); 245 optimizer.addParameter(camera); 246 247 // edges 248 index = 1; 249 for (const Point2f p : points_2d) 250 { 251 g2o::EdgeProjectXYZ2UV* edge = new g2o::EdgeProjectXYZ2UV(); 252 edge->setId(index); 253 edge->setVertex(0, dynamic_cast<g2o::VertexSBAPointXYZ*> (optimizer.vertex(index))); 254 edge->setVertex(1, pose); 255 edge->setMeasurement(Eigen::Vector2d(p.x, p.y)); 256 edge->setParameterId(0, 0); 257 edge->setInformation(Eigen::Matrix2d::Identity()); 258 optimizer.addEdge(edge); 259 index++; 260 } 261 262 chrono::steady_clock::time_point t1 = chrono::steady_clock::now(); 263 optimizer.setVerbose(true); 264 optimizer.initializeOptimization(); 265 optimizer.optimize(100); 266 chrono::steady_clock::time_point t2 = chrono::steady_clock::now(); 267 chrono::duration<double> time_used = chrono::duration_cast<chrono::duration<double>> (t2 - t1); 268 cout << "optimization costs time: " << time_used.count() << " seconds." << endl; 269 270 cout << endl << "after optimization:" << endl; 271 cout << "T=" << endl << Eigen::Isometry3d(pose->estimate()).matrix() << endl; 272 }
這里我們只關心運動,不關心結構。換句話說,只要通過特征點成功求出了運動,我們就不再需要這幀的特征點了。這種做法當然會有缺陷,但是忽略掉數量龐大的特征點可以節省許多的計算量。然后,在 t 到 t + 1 時刻,我們又以 t 時刻為參考幀,考慮 t 到 t + 1 間的運動關系。如此往復,就得到了一條運動軌跡。