在之前的章节我已经讲过了G2O(General Graph Optimization)的相关原理以及应用,这次的BA优化问题呢,我们也可以转化为一个G2O来求解。
G2O(General Graph Optimization)———— 通用图优化。
G2O(General Graph Optimization)的核里带有种类多样的求解器,而它的顶点、边的类型也是多种多样。我们可以自己定义顶点和边。总的来说,如果一个优化问题能够表达成图(顶点与边),那么这个问题就可以用G2O(General Graph Optimization)去求解它。常见的,比如bundle adjustment(这里的BA优化),ICP,数据拟合,都可以用G2O(General Graph Optimization)来做。
如何使用图优化进行BA求解呢?
对于图优化问题,首先需要定义一个顶点和一个边。这里呢,我们使用顶点为第二个相机帧的位姿作为顶点,即节点。对于边呢,就是定义相机种的三维点在第二个相机帧的投影作为边。
void find_feature_matches(
const Mat &img_1, const Mat &img_2,
std::vector<keypoint> &keypoints_1,
std::vector<keypoint> &keypoints_2,
std::vector<dmatch> &matches);
// 像素坐标转相机归一化坐标
Point2d pixel2cam(const Point2d &p, const Mat &K);
// BA by g2o
typedef vector<eigen::vector2d, eigen::aligned_allocator<eigen::vector2d="">> VecVector2d;
typedef vector<eigen::vector3d, eigen::aligned_allocator<eigen::vector3d="">> VecVector3d;
void bundleAdjustmentG2O(
const VecVector3d &points_3d,
const VecVector2d &points_2d,
const Mat &K,
Sophus::SE3d &pose
);
// BA by gauss-newton
void bundleAdjustmentGaussNewton(
const VecVector3d &points_3d,
const VecVector2d &points_2d,
const Mat &K,
Sophus::SE3d &pose
);
int main(int argc, char **argv) {
if (argc != 5) {
cout << "usage: pose_estimation_3d2d img1 img2 depth1 depth2" << endl;
return 1;
}
//-- 读取图像
Mat img_1 = imread(argv[1], CV_LOAD_IMAGE_COLOR);
Mat img_2 = imread(argv[2], CV_LOAD_IMAGE_COLOR);
assert(img_1.data && img_2.data && "Can not load images!");
vector<keypoint> keypoints_1, keypoints_2;
vector<dmatch> matches;
find_feature_matches(img_1, img_2, keypoints_1, keypoints_2, matches);
cout << "一共找到了" << matches.size() << "组匹配点" << endl;
// 建立3D点
Mat d1 = imread(argv[3], CV_LOAD_IMAGE_UNCHANGED); // 深度图为16位无符号数,单通道图像
Mat K = (Mat_<double>(3, 3) << 520.9, 0, 325.1, 0, 521.0, 249.7, 0, 0, 1);
vector<point3f> pts_3d;
vector<point2f> pts_2d;
for (DMatch m:matches) {
ushort d = d1.ptr<unsigned short="">(int(keypoints_1[m.queryIdx].pt.y))[int(keypoints_1[m.queryIdx].pt.x)];
if (d == 0) // bad depth
continue;
float dd = d / 5000.0;
Point2d p1 = pixel2cam(keypoints_1[m.queryIdx].pt, K);
pts_3d.push_back(Point3f(p1.x * dd, p1.y * dd, dd));
pts_2d.push_back(keypoints_2[m.trainIdx].pt);
}
cout << "3d-2d pairs: " << pts_3d.size() << endl;
chrono::steady_clock::time_point t1 = chrono::steady_clock::now();
Mat r, t;
solvePnP(pts_3d, pts_2d, K, Mat(), r, t, false); // 调用OpenCV 的 PnP 求解,可选择EPNP,DLS等方法
Mat R;
cv::Rodrigues(r, R); // r为旋转向量形式,用Rodrigues公式转换为矩阵
chrono::steady_clock::time_point t2 = chrono::steady_clock::now();
chrono::duration<double> time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);
cout << "solve pnp in opencv cost time: " << time_used.count() << " seconds." << endl;
cout << "R=" << endl << R << endl;
cout << "t=" << endl << t << endl;
VecVector3d pts_3d_eigen;
VecVector2d pts_2d_eigen;
for (size_t i = 0; i < pts_3d.size(); ++i) {
pts_3d_eigen.push_back(Eigen::Vector3d(pts_3d[i].x, pts_3d[i].y, pts_3d[i].z));
pts_2d_eigen.push_back(Eigen::Vector2d(pts_2d[i].x, pts_2d[i].y));
}
cout << "calling bundle adjustment by gauss newton" << endl;
Sophus::SE3d pose_gn;
t1 = chrono::steady_clock::now();
bundleAdjustmentGaussNewton(pts_3d_eigen, pts_2d_eigen, K, pose_gn);
t2 = chrono::steady_clock::now();
time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);
cout << "solve pnp by gauss newton cost time: " << time_used.count() << " seconds." << endl;
cout << "calling bundle adjustment by g2o" << endl;
Sophus::SE3d pose_g2o;
t1 = chrono::steady_clock::now();
bundleAdjustmentG2O(pts_3d_eigen, pts_2d_eigen, K, pose_g2o);
t2 = chrono::steady_clock::now();
time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);
cout << "solve pnp by g2o cost time: " << time_used.count() << " seconds." << endl;
return 0;
}
Point2d pixel2cam(const Point2d &p, const Mat &K) {
return Point2d
(
(p.x - K.at<double>(0, 2)) / K.at<double>(0, 0),
(p.y - K.at<double>(1, 2)) / K.at<double>(1, 1)
);
}
void bundleAdjustmentGaussNewton(
const VecVector3d &points_3d,
const VecVector2d &points_2d,
const Mat &K,
Sophus::SE3d &pose) {
typedef Eigen::Matrix<double, 6,="" 1=""> Vector6d;
const int iterations = 10;
double cost = 0, lastCost = 0;
double fx = K.at<double>(0, 0);
double fy = K.at<double>(1, 1);
double cx = K.at<double>(0, 2);
double cy = K.at<double>(1, 2);
for (int iter = 0; iter < iterations; iter++) {
Eigen::Matrix<double, 6,="" 6=""> H = Eigen::Matrix<double, 6,="" 6="">::Zero();
Vector6d b = Vector6d::Zero();
cost = 0;
// compute cost
for (int i = 0; i < points_3d.size(); i++) {
Eigen::Vector3d pc = pose * points_3d[i];
double inv_z = 1.0 / pc[2];
double inv_z2 = inv_z * inv_z;
Eigen::Vector2d proj(fx * pc[0] / pc[2] + cx, fy * pc[1] / pc[2] + cy);
Eigen::Vector2d e = points_2d[i] - proj;
cost += e.squaredNorm();
Eigen::Matrix<double, 2,="" 6=""> J;
J << -fx * inv_z,
0,
fx * pc[0] * inv_z2,
fx * pc[0] * pc[1] * inv_z2,
-fx - fx * pc[0] * pc[0] * inv_z2,
fx * pc[1] * inv_z,
0,
-fy * inv_z,
fy * pc[1] * inv_z2,
fy + fy * pc[1] * pc[1] * inv_z2,
-fy * pc[0] * pc[1] * inv_z2,
-fy * pc[0] * inv_z;
H += J.transpose() * J;
b += -J.transpose() * e;
}
Vector6d dx;
dx = H.ldlt().solve(b);
if (isnan(dx[0])) {
cout << "result is nan!" << endl;
break;
}
if (iter > 0 && cost >= lastCost) {
// cost increase, update is not good
cout << "cost: " << cost << ", last cost: " << lastCost << endl;
break;
}
// precision(12) << cost << endl;
if (dx.norm() < 1e-6) {
// converge
break;
}
}
cout << "pose by g-n: \n" << pose.matrix() << endl;
}
/// vertex and edges used in g2o ba
class VertexPose : public g2o::BaseVertex<6, Sophus::SE3d> {
public:
EIGEN_MAKE_ALIGNED_OPERATOR_NEW;
virtual void setToOriginImpl() override {
_estimate = Sophus::SE3d();
}
/// left multiplication on SE3
virtual void oplusImpl(const double *update) override {
Eigen::Matrix<double, 6,="" 1=""> update_eigen;
update_eigen << update[0], update[1], update[2], update[3], update[4], update[5];
_estimate = Sophus::SE3d::exp(update_eigen) * _estimate;
}
virtual bool read(istream &in) override {}
virtual bool write(ostream &out) const override {}
};
class EdgeProjection : public g2o::BaseUnaryEdge<2, Eigen::Vector2d, VertexPose> {
public:
EIGEN_MAKE_ALIGNED_OPERATOR_NEW;
EdgeProjection(const Eigen::Vector3d &pos, const Eigen::Matrix3d &K) : _pos3d(pos), _K(K) {}
virtual void computeError() override {
const VertexPose *v = static_cast<vertexpose *=""> (_vertices[0]);
Sophus::SE3d T = v->estimate();
Eigen::Vector3d pos_pixel = _K * (T * _pos3d);
pos_pixel /= pos_pixel[2];
_error = _measurement - pos_pixel.head<2>();
}
virtual void linearizeOplus() override {
const VertexPose *v = static_cast<vertexpose *=""> (_vertices[0]);
Sophus::SE3d T = v->estimate();
Eigen::Vector3d pos_cam = T * _pos3d;
double fx = _K(0, 0);
double fy = _K(1, 1);
double cx = _K(0, 2);
double cy = _K(1, 2);
double X = pos_cam[0];
double Y = pos_cam[1];
double Z = pos_cam[2];
double Z2 = Z * Z;
_jacobianOplusXi
<< -fx / Z, 0, fx * X / Z2, fx * X * Y / Z2, -fx - fx * X * X / Z2, fx * Y / Z,
0, -fy / Z, fy * Y / (Z * Z), fy + fy * Y * Y / Z2, -fy * X * Y / Z2, -fy * X / Z;
}
virtual bool read(istream &in) override {}
virtual bool write(ostream &out) const override {}
private:
Eigen::Vector3d _pos3d;
Eigen::Matrix3d _K;
};
void bundleAdjustmentG2O(
const VecVector3d &points_3d,
const VecVector2d &points_2d,
const Mat &K,
Sophus::SE3d &pose) {
// 构建图优化,先设定g2o
typedef g2o::BlockSolver<g2o::blocksolvertraits<6, 3="">> BlockSolverType; // pose is 6, landmark is 3
typedef g2o::LinearSolverDense<blocksolvertype::posematrixtype> LinearSolverType; // 线性求解器类型
// 梯度下降方法,可以从GN, LM, DogLeg 中选
auto solver = new g2o::OptimizationAlgorithmGaussNewton(
g2o::make_unique<blocksolvertype>(g2o::make_unique<linearsolvertype>()));
g2o::SparseOptimizer optimizer; // 图模型
optimizer.setAlgorithm(solver); // 设置求解器
optimizer.setVerbose(true); // 打开调试输出
// vertex
VertexPose *vertex_pose = new VertexPose(); // camera vertex_pose
vertex_pose->setId(0);
vertex_pose->setEstimate(Sophus::SE3d());
optimizer.addVertex(vertex_pose);
// K
Eigen::Matrix3d K_eigen;
K_eigen <<
K.at<double>(0, 0), K.at<double>(0, 1), K.at<double>(0, 2),
K.at<double>(1, 0), K.at<double>(1, 1), K.at<double>(1, 2),
K.at<double>(2, 0), K.at<double>(2, 1), K.at<double>(2, 2);
// edges
int index = 1;
for (size_t i = 0; i < points_2d.size(); ++i) {
auto p2d = points_2d[i];
auto p3d = points_3d[i];
EdgeProjection *edge = new EdgeProjection(p3d, K_eigen);
edge->setId(index);
edge->setVertex(0, vertex_pose);
edge->setMeasurement(p2d);
edge->setInformation(Eigen::Matrix2d::Identity());
optimizer.addEdge(edge);
index++;
}
chrono::steady_clock::time_point t1 = chrono::steady_clock::now();
optimizer.setVerbose(true);
optimizer.initializeOptimization();
optimizer.optimize(10);
chrono::steady_clock::time_point t2 = chrono::steady_clock::now();
chrono::duration<double> time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);
cout << "optimization costs time: " << time_used.count() << " seconds." << endl;
cout << "pose estimated by g2o =\n" << vertex_pose->estimate().matrix() << endl;
pose = vertex_pose->estimate();
}
ICP(Iterative Closest Point)———— 迭代最近点。这是一个3D-3D点的优化问题,具体就是在三维点知道了如何匹配特征点,通过三维特征点匹配出相关的位姿优化问题。所以ICP(Iterative Closest Point)可以是经常在激光雷达的SLAM构建中经常遇到。
对于ICP(Iterative Closest Point)问题的求解,在学术领域主要分为2种类型:第一:使用奇异值分解求解;第二:使用非线性优化求解。
> ICP算法流程
首先对于一幅点云中的每个点,在另一幅点云中计算匹配点(最近点)
极小化匹配点间的匹配误差,计算位姿
然后将计算的位姿作用于点云
再重新计算匹配点
如此迭代,直到迭代次数达到阈值,或者极小化的能量函数变化量小于设定阈值
对于使用奇异值求解的办法,即SVD方法。首先对于已经知道的2个三维点进行投影误差的计算,构建最小二乘问题。分别求取旋转矩阵,在求取平移矩阵。
我们依然构建三维点误差的最小二乘估计,通过之前介绍的PNP类似的办法求取优化。如果使用李群李代数,则在求导时用李群李代数的扰动模型。
/// g2o edge
class EdgeProjectXYZRGBDPoseOnly : public g2o::BaseUnaryEdge<3, Eigen::Vector3d, VertexPose> {
public:
EIGEN_MAKE_ALIGNED_OPERATOR_NEW;
EdgeProjectXYZRGBDPoseOnly(const Eigen::Vector3d &point) : _point(point) {}
virtual void computeError() override {
const VertexPose *pose = static_cast<const vertexpose="" *=""> ( _vertices[0] );
_error = _measurement - pose->estimate() * _point;
}
virtual void linearizeOplus() override {
VertexPose *pose = static_cast<vertexpose *="">(_vertices[0]);
Sophus::SE3d T = pose->estimate();
Eigen::Vector3d xyz_trans = T * _point;
_jacobianOplusXi.block<3, 3>(0, 0) = -Eigen::Matrix3d::Identity();
_jacobianOplusXi.block<3, 3>(0, 3) = Sophus::SO3d::hat(xyz_trans);
}
bool read(istream &in) {}
bool write(ostream &out) const {}
protected:
Eigen::Vector3d _point;
};
int main(int argc, char **argv) {
vector<keypoint> keypoints_1, keypoints_2;
vector<dmatch> matches;
find_feature_matches(img_1, img_2, keypoints_1, keypoints_2, matches);
cout << "一共找到了" << matches.size() << "组匹配点" << endl;
// 建立3D点
Mat depth1 = imread(argv[3], CV_LOAD_IMAGE_UNCHANGED); // 深度图为16位无符号数,单通道图像
Mat depth2 = imread(argv[4], CV_LOAD_IMAGE_UNCHANGED); // 深度图为16位无符号数,单通道图像
Mat K = (Mat_<double>(3, 3) << 520.9, 0, 325.1, 0, 521.0, 249.7, 0, 0, 1);
vector<point3f> pts1, pts2;
for (DMatch m:matches) {
ushort d1 = depth1.ptr<unsigned short="">(int(keypoints_1[m.queryIdx].pt.y))[int(keypoints_1[m.queryIdx].pt.x)];
ushort d2 = depth2.ptr<unsigned short="">(int(keypoints_2[m.trainIdx].pt.y))[int(keypoints_2[m.trainIdx].pt.x)];
}
cout << "3d-3d pairs: " << pts1.size() << endl;
Mat R, t;
pose_estimation_3d3d(pts1, pts2, R, t);
cout << "calling bundle adjustment" << endl;
bundleAdjustment(pts1, pts2, R, t);
}
}
void find_feature_matches(const Mat &img_1, const Mat &img_2,
std::vector<keypoint> &keypoints_1,
std::vector<keypoint> &keypoints_2,
std::vector<dmatch> &matches) {
//-- 初始化
Mat descriptors_1, descriptors_2;
// used in OpenCV3
Ptr<featuredetector> detector = ORB::create();
Ptr<descriptorextractor> descriptor = ORB::create();
// use this if you are in OpenCV2
// Ptr detector = FeatureDetector::create ( "ORB" );
// Ptr descriptor = DescriptorExtractor::create ( "ORB" );
Ptr<descriptormatcher> matcher = DescriptorMatcher::create("BruteForce-Hamming");
//-- 第一步:检测 Oriented FAST 角点位置
detector->detect(img_1, keypoints_1);
detector->detect(img_2, keypoints_2);
//-- 第二步:根据角点位置计算 BRIEF 描述子
descriptor->compute(img_1, keypoints_1, descriptors_1);
descriptor->compute(img_2, keypoints_2, descriptors_2);
//-- 第三步:对两幅图像中的BRIEF描述子进行匹配,使用 Hamming 距离
vector<dmatch> match;
// BFMatcher matcher ( NORM_HAMMING );
matcher->match(descriptors_1, descriptors_2, match);
//-- 第四步:匹配点对筛选
double min_dist = 10000, max_dist = 0;
//找出所有匹配之间的最小距离和最大距离, 即是最相似的和最不相似的两组点之间的距离
for (int i = 0; i < descriptors_1.rows; i++) {
double dist = match[i].distance;
if (dist < min_dist) min_dist = dist;
if (dist > max_dist) max_dist = dist;
}
printf("-- Max dist : %f \n", max_dist);
printf("-- Min dist : %f \n", min_dist);
//当描述子之间的距离大于两倍的最小距离时,即认为匹配有误.但有时候最小距离会非常小,设置一个经验值30作为下限.
for (int i = 0; i < descriptors_1.rows; i++) {
if (match[i].distance <= max(2 * min_dist, 30.0)) {
matches.push_back(match[i]);
}
}
}
void pose_estimation_3d3d(const vector<point3f> &pts1,
const vector<point3f> &pts2,
Mat &R, Mat &t) {
Point3f p1, p2; // center of mass
int N = pts1.size();
for (int i = 0; i < N; i++) {
p1 += pts1[i];
p2 += pts2[i];
}
p1 = Point3f(Vec3f(p1) / N);
p2 = Point3f(Vec3f(p2) / N);
vector<point3f> q1(N), q2(N); // remove the center
for (int i = 0; i < N; i++) {
q1[i] = pts1[i] - p1;
q2[i] = pts2[i] - p2;
}
// compute q1*q2^T
Eigen::Matrix3d W = Eigen::Matrix3d::Zero();
for (int i = 0; i < N; i++) {
W += Eigen::Vector3d(q1[i].x, q1[i].y, q1[i].z) * Eigen::Vector3d(q2[i].x, q2[i].y, q2[i].z).transpose();
}
cout << "W=" << W << endl;
// SVD on W
Eigen::JacobiSVD<eigen::matrix3d> svd(W, Eigen::ComputeFullU | Eigen::ComputeFullV);
Eigen::Matrix3d U = svd.matrixU();
Eigen::Matrix3d V = svd.matrixV();
cout << "U=" << U << endl;
cout << "V=" << V << endl;
Eigen::Matrix3d R_ = U * (V.transpose());
if (R_.determinant() < 0) {
R_ = -R_;
}
Eigen::Vector3d t_ = Eigen::Vector3d(p1.x, p1.y, p1.z) - R_ * Eigen::Vector3d(p2.x, p2.y, p2.z);
// convert to cv::Mat
R = (Mat_<double>(3, 3) <<
R_(0, 0), R_(0, 1), R_(0, 2),
R_(1, 0), R_(1, 1), R_(1, 2),
R_(2, 0), R_(2, 1), R_(2, 2)
);
t = (Mat_<double>(3, 1) << t_(0, 0), t_(1, 0), t_(2, 0));
}
void bundleAdjustment(
const vector<point3f> &pts1,
const vector<point3f> &pts2,
Mat &R, Mat &t) {
// 构建图优化,先设定g2o
typedef g2o::BlockSolverX BlockSolverType;
typedef g2o::LinearSolverDense<blocksolvertype::posematrixtype> LinearSolverType; // 线性求解器类型
// 梯度下降方法,可以从GN, LM, DogLeg 中选
auto solver = new g2o::OptimizationAlgorithmLevenberg(
g2o::make_unique<blocksolvertype>(g2o::make_unique<linearsolvertype>()));
g2o::SparseOptimizer optimizer; // 图模型
optimizer.setAlgorithm(solver); // 设置求解器
optimizer.setVerbose(true); // 打开调试输出
// vertex
VertexPose *pose = new VertexPose(); // camera pose
pose->setId(0);
pose->setEstimate(Sophus::SE3d());
optimizer.addVertex(pose);
// edges
for (size_t i = 0; i < pts1.size(); i++) {
EdgeProjectXYZRGBDPoseOnly *edge = new EdgeProjectXYZRGBDPoseOnly(
Eigen::Vector3d(pts2[i].x, pts2[i].y, pts2[i].z));
edge->setVertex(0, pose);
edge->setMeasurement(Eigen::Vector3d(
pts1[i].x, pts1[i].y, pts1[i].z));
edge->setInformation(Eigen::Matrix3d::Identity());
optimizer.addEdge(edge);
}
chrono::steady_clock::time_point t1 = chrono::steady_clock::now();
optimizer.initializeOptimization();
optimizer.optimize(10);
chrono::steady_clock::time_point t2 = chrono::steady_clock::now();
chrono::duration<double> time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);
cout << "optimization costs time: " << time_used.count() << " seconds." << endl;
cout << endl << "after optimization:" << endl;
cout << "T=\n" << pose->estimate().matrix() << endl;
// convert to cv::Mat
Eigen::Matrix3d R_ = pose->estimate().rotationMatrix();
Eigen::Vector3d t_ = pose->estimate().translation();
R = (Mat_<double>(3, 3) <<
R_(0, 0), R_(0, 1), R_(0, 2),
R_(1, 0), R_(1, 1), R_(1, 2),
R_(2, 0), R_(2, 1), R_(2, 2)
);
t = (Mat_<double>(3, 1) << t_(0, 0), t_(1, 0), t_(2, 0));
}
到这里,我们已经介绍完了2D-2D,3D-2D(PNP),3D-3D问题的求解,对于前端的位姿估计想必也有了更深的理解,这里其实还有很多问题,在实际运行的机器人上还会存在各种各样的情况与误差,我在未来会继续讲解这一部分的问题和相关求解。
> 参考资料:视觉SLAM十四讲