前言
在SfM多視圖三維點雲重建–【VS2015+OpenCV3.4+PCL1.8】中實現的增量式SfM三維點雲重建,會隨着圖片數量的增加而導致誤差逐漸累積,最後可能無法完成重建。在三維重建中常使用Bundle Adjustment(BA)優化來解決這個問題,BA優化是一種應用廣泛的非線性優化方法,其本質屬於非線性最小二乘問題,目的是找到使得重投影誤差最小的3D點位置和相機參數。這裏對BA原理就不再贅述,而是使用Ceres Solver對SfM三維重建後的內外參、點雲進行BA優化。
代碼
環境:VS2015+OpenCV3.4+PCL1.8+Ceres Solver
首先定義重投影誤差作爲待優化的代價函數,即計算三維空間點的重投影(世界座標系到像素座標系),並計算重投影誤差,這部分有很多種形式,但都大同小異。這裏ceres::AngleAxisRotatePoint(r, pos3d, pos_proj)將空間點pos3d進行旋轉變換(旋轉向量r)得到pos_proj,另外,注意添加如下定義和頭文件,否則會報錯。
#ifndef GLOG_NO_ABBREVIATED_SEVERITIES
#define GLOG_NO_ABBREVIATED_SEVERITIES
#endif
#include <ceres/ceres.h>
#include <ceres/rotation.h>
// 代價函數
struct ReprojectCost
{
cv::Point2d observation;
ReprojectCost(cv::Point2d& observation) : observation(observation){}
// 參數:內參、外參、三維點、反向投影誤差
template <typename T>
bool operator()(const T* const intrinsic, const T* const extrinsic, const T* const pos3d, T* residuals) const
{
const T* r = extrinsic;
const T* t = &extrinsic[3];
// 外參數矩陣:轉到相機座標系
// Apply the camera rotation
T pos_proj[3];
ceres::AngleAxisRotatePoint(r, pos3d, pos_proj);
// Apply the camera translation
pos_proj[0] += t[0];
pos_proj[1] += t[1];
pos_proj[2] += t[2];
const T x = pos_proj[0] / pos_proj[2]; // 歸一化相機座標
const T y = pos_proj[1] / pos_proj[2];
// 內參數矩陣:轉到像素座標系
const T fx = intrinsic[0];
const T fy = intrinsic[1];
const T cx = intrinsic[2];
const T cy = intrinsic[3];
// Apply intrinsic
const T u = fx * x + cx;
const T v = fy * y + cy;
// 計算重投影誤差
residuals[0] = u - T(observation.x);
residuals[1] = v - T(observation.y);
return true;
}
};
接下來使用Ceres Solver求解BA,前兩張圖片的BA優化函數如下:
/*
main函數中添加:
// BA優化
Mat intrinsic(Matx41d(K.at<double>(0, 0), K.at<double>(1, 1), K.at<double>(0, 2), K.at<double>(1, 2)));
Mat extrinsic(6, 1, CV_64FC1);
Mat r;
Rodrigues(R, r);
r.copyTo(extrinsic.rowRange(0, 3));// [0,3)
t.copyTo(extrinsic.rowRange(3, 6));
bundle_adjustment(intrinsic, extrinsic, points1, points3D);
*/
void bundle_adjustment(Mat& intrinsic, Mat& extrinsics, vector<Point2d>& points, vector<Point3d>& points3D)
{
ceres::Problem problem;
// load extrinsics (rotations and motions)
problem.AddParameterBlock(extrinsics.ptr<double>(), 6);
// fix the first camera.
problem.SetParameterBlockConstant(extrinsics.ptr<double>());
// load intrinsic
problem.AddParameterBlock(intrinsic.ptr<double>(), 4); // fx, fy, cx, cy
// load points
ceres::LossFunction* loss_function = new ceres::HuberLoss(4); // loss function make bundle adjustment robuster.
for (size_t i = 0; i < points3D.size(); i++)
{
Point2d observed = points[i];
// 模板參數:代價函數的類型、代價的維度、代價函數第一、第二、第三個參數的維度
ceres::CostFunction* cost_function = new ceres::AutoDiffCostFunction<ReprojectCost, 2, 4, 6, 3>(new ReprojectCost(observed));
problem.AddResidualBlock(
cost_function,
loss_function,
intrinsic.ptr<double>(), // Intrinsic
extrinsics.ptr<double>(), // View Rotation and Translation
&(points3D[i].x) // Point in 3D space
);
}
// Solve BA
ceres::Solver::Options ceres_config_options;
ceres_config_options.minimizer_progress_to_stdout = false;
ceres_config_options.logging_type = ceres::SILENT;
ceres_config_options.num_threads = 1;
ceres_config_options.preconditioner_type = ceres::JACOBI;
ceres_config_options.linear_solver_type = ceres::SPARSE_SCHUR;
ceres_config_options.sparse_linear_algebra_library_type = ceres::EIGEN_SPARSE;
ceres::Solver::Summary summary;
ceres::Solve(ceres_config_options, &problem, &summary);
if (!summary.IsSolutionUsable())
{
std::cout << "Bundle Adjustment failed." << std::endl;
}
else
{
// Display statistics about the minimization
std::cout << std::endl
<< "Bundle Adjustment statistics (approximated RMSE):\n"
<< " #views: 0\n"
<< " #num_residuals: " << summary.num_residuals << "\n"
<< " Initial RMSE: " << std::sqrt(summary.initial_cost / summary.num_residuals) << "\n"
<< " Final RMSE: " << std::sqrt(summary.final_cost / summary.num_residuals) << "\n"
<< " Time (s): " << summary.total_time_in_seconds << "\n"
<< std::endl;
}
}
如下兩圖是前兩張圖片BA優化前後的結果,雖然直觀上來看,兩者並沒有什麼太大的區別。但是根據BA優化後的統計信息可知,BA優化前重投影的三維點與真實點平均偏移11.68個像素,優化後兩者幾乎沒有偏差(RMSE = 0),可見還是有了很大的改善。
多張圖片BA優化函數:
/*
main函數中添加:
// BA優化
Mat intrinsic(Matx41d(K.at<double>(0, 0), K.at<double>(1, 1), K.at<double>(0, 2), K.at<double>(1, 2)));
vector<Mat> extrinsics;
for (size_t i = 0; i < rotations.size(); ++i)
{
Mat extrinsic(6, 1, CV_64FC1);
Mat r;
Rodrigues(rotations[i], r);
r.copyTo(extrinsic.rowRange(0, 3));
translations[i].copyTo(extrinsic.rowRange(3, 6));
extrinsics.push_back(extrinsic);
}
bundle_adjustment(intrinsic, extrinsics, correspond_struct_idx, key_points_for_all, points3D);
*/
void bundle_adjustment(Mat& intrinsic, vector<Mat>& extrinsics, vector<vector<int>>& correspond_struct_idx, vector<vector<KeyPoint>>& key_points_for_all, vector<Point3d>& structure)
{
// 構建問題,求解
ceres::Problem problem;
// load extrinsics
for (size_t i = 0; i < extrinsics.size(); ++i)
{
problem.AddParameterBlock(extrinsics[i].ptr<double>(), 6);
}
// load intrinsic
problem.AddParameterBlock(intrinsic.ptr<double>(), 4); // fx, fy, cx, cy
// fix the first camera.
problem.SetParameterBlockConstant(extrinsics[0].ptr<double>());
// 添加初始點
ceres::LossFunction* loss_function = new ceres::HuberLoss(4); // loss function make bundle adjustment robuster.
for (size_t img_idx = 0; img_idx < correspond_struct_idx.size(); ++img_idx)
{
vector<int>& point3d_ids = correspond_struct_idx[img_idx];
vector<KeyPoint>& key_points = key_points_for_all[img_idx];
for (size_t point_idx = 0; point_idx < point3d_ids.size(); ++point_idx)
{
int point3d_id = point3d_ids[point_idx];
if (point3d_id < 0)
continue;
Point2d observed = key_points[point_idx].pt;
// 模板參數:代價函數的類型、代價的維度、代價函數第一、第二、第三個參數的維度
ceres::CostFunction* cost_function = new ceres::AutoDiffCostFunction<ReprojectCost, 2, 4, 6, 3>(new ReprojectCost(observed));
problem.AddResidualBlock(
cost_function,
loss_function,
intrinsic.ptr<double>(), // Intrinsic
extrinsics[img_idx].ptr<double>(), // View Rotation and Translation
&(structure[point3d_id].x) // Point in 3D space
);
}
}
// 設置優化策略
ceres::Solver::Options ceres_config_options;
ceres_config_options.minimizer_progress_to_stdout = false;
ceres_config_options.logging_type = ceres::SILENT;
ceres_config_options.num_threads = 1;
ceres_config_options.preconditioner_type = ceres::JACOBI;
ceres_config_options.linear_solver_type = ceres::SPARSE_SCHUR;
ceres_config_options.sparse_linear_algebra_library_type = ceres::EIGEN_SPARSE;
// 求解
ceres::Solver::Summary summary;
ceres::Solve(ceres_config_options, &problem, &summary);
// 打印結果
if (!summary.IsSolutionUsable())
{
std::cout << "Bundle Adjustment failed." << std::endl;
}
else
{
// Display statistics about the minimization
std::cout << std::endl
<< "Bundle Adjustment statistics (approximated RMSE):\n"
<< " #views: " << extrinsics.size() << "\n"
<< " #num_residuals: " << summary.num_residuals << "\n"
<< " Initial RMSE: " << std::sqrt(summary.initial_cost / summary.num_residuals) << "\n"
<< " Final RMSE: " << std::sqrt(summary.final_cost / summary.num_residuals) << "\n"
<< " Time (s): " << summary.total_time_in_seconds << "\n"
<< std::endl;
}
}
多張圖片BA優化結果: