electech6 2021-08-09 14:16:10 +08:00
commit 43c195710b
3 changed files with 424 additions and 43 deletions

View File

@ -1,7 +1,7 @@
{ {
"configurations": [ "configurations": [
{ {
"name": "Win32", "name": "Linux",
"includePath": [ "includePath": [
"${workspaceFolder}/**" "${workspaceFolder}/**"
], ],
@ -11,7 +11,7 @@
"_UNICODE" "_UNICODE"
], ],
"windowsSdkVersion": "10.0.18362.0", "windowsSdkVersion": "10.0.18362.0",
"compilerPath": "d:/Program Files (x86)/Microsoft Visual Studio/2019/Community/VC/Tools/MSVC/14.24.28314/bin/Hostx64/x64/cl.exe", "compilerPath": "/usr/bin/g++",
"cStandard": "c11", "cStandard": "c11",
"cppStandard": "c++17", "cppStandard": "c++17",
"intelliSenseMode": "msvc-x64" "intelliSenseMode": "msvc-x64"

View File

@ -51,40 +51,68 @@
namespace ORB_SLAM3 { namespace ORB_SLAM3 {
/**
* @brief MLPnP
*
* @param[in] F
* @param[in] vpMapPointMatches
* @param[in] mnInliersi
* @param[in] mnIterations Ransac
* @param[in] mnBestInliers
* @param[in] N 2D
* @param[in] mpCamera 3D
*/
MLPnPsolver::MLPnPsolver(const Frame &F, const vector<MapPoint *> &vpMapPointMatches): MLPnPsolver::MLPnPsolver(const Frame &F, const vector<MapPoint *> &vpMapPointMatches):
mnInliersi(0), mnIterations(0), mnBestInliers(0), N(0), mpCamera(F.mpCamera){ mnInliersi(0), // 内点的个数
mvpMapPointMatches = vpMapPointMatches; mnIterations(0), // Ransac迭代次数
mvBearingVecs.reserve(F.mvpMapPoints.size()); mnBestInliers(0), // 最佳内点数
mvP2D.reserve(F.mvpMapPoints.size()); N(0), // 所有2D点的个数
mvSigma2.reserve(F.mvpMapPoints.size()); mpCamera(F.mpCamera) // 相机模型利用该变量对3D点进行投影
mvP3Dw.reserve(F.mvpMapPoints.size()); {
mvKeyPointIndices.reserve(F.mvpMapPoints.size()); mvpMapPointMatches = vpMapPointMatches; // 待匹配的特征点是当前帧和候选关键帧用BoW进行快速匹配的结果
mvAllIndices.reserve(F.mvpMapPoints.size()); mvBearingVecs.reserve(F.mvpMapPoints.size()); // ? 初始化3D点的单位向量
mvP2D.reserve(F.mvpMapPoints.size()); // 初始化3D点的投影点
mvSigma2.reserve(F.mvpMapPoints.size()); // 初始化卡方检验中的sigma值
mvP3Dw.reserve(F.mvpMapPoints.size()); // 初始化3D点坐标
mvKeyPointIndices.reserve(F.mvpMapPoints.size()); // 初始化3D点的索引值
mvAllIndices.reserve(F.mvpMapPoints.size()); // 初始化所有索引值
// 一些必要的初始化操作
int idx = 0; int idx = 0;
for(size_t i = 0, iend = mvpMapPointMatches.size(); i < iend; i++){ for(size_t i = 0, iend = mvpMapPointMatches.size(); i < iend; i++){
MapPoint* pMP = vpMapPointMatches[i]; MapPoint* pMP = vpMapPointMatches[i];
// 如果pMP存在则接下来初始化一些参数否则什么都不做
if(pMP){ if(pMP){
// 判断是否是坏点
if(!pMP -> isBad()){ if(!pMP -> isBad()){
// 如果记录的点个数超过总数,则不做任何事情,否则继续记录
if(i >= F.mvKeysUn.size()) continue; if(i >= F.mvKeysUn.size()) continue;
const cv::KeyPoint &kp = F.mvKeysUn[i]; const cv::KeyPoint &kp = F.mvKeysUn[i];
// 保存3D点的投影点
mvP2D.push_back(kp.pt); mvP2D.push_back(kp.pt);
// 保存卡方检验中的sigma值
mvSigma2.push_back(F.mvLevelSigma2[kp.octave]); mvSigma2.push_back(F.mvLevelSigma2[kp.octave]);
//Bearing vector should be normalized //Bearing vector should be normalized
// 特征点投影,并计算单位向量
cv::Point3f cv_br = mpCamera->unproject(kp.pt); cv::Point3f cv_br = mpCamera->unproject(kp.pt);
cv_br /= cv_br.z; cv_br /= cv_br.z;
bearingVector_t br(cv_br.x,cv_br.y,cv_br.z); bearingVector_t br(cv_br.x,cv_br.y,cv_br.z);
mvBearingVecs.push_back(br); mvBearingVecs.push_back(br);
//3D coordinates //3D coordinates
// 获取当前特征点的3D坐标
cv::Mat cv_pos = pMP -> GetWorldPos(); cv::Mat cv_pos = pMP -> GetWorldPos();
point_t pos(cv_pos.at<float>(0),cv_pos.at<float>(1),cv_pos.at<float>(2)); point_t pos(cv_pos.at<float>(0),cv_pos.at<float>(1),cv_pos.at<float>(2));
mvP3Dw.push_back(pos); mvP3Dw.push_back(pos);
// 记录当前特征点的索引值,挑选后的
mvKeyPointIndices.push_back(i); mvKeyPointIndices.push_back(i);
// 记录所有特征点的索引值
mvAllIndices.push_back(idx); mvAllIndices.push_back(idx);
idx++; idx++;
@ -96,57 +124,89 @@ namespace ORB_SLAM3 {
} }
//RANSAC methods //RANSAC methods
/**
* @brief MLPnP姿
*
* @param[in] nIterations
* @param[in] bNoMore
* @param[in] vbInliers
* @param[in] nInliers
* @return cv::Mat 姿
*/
cv::Mat MLPnPsolver::iterate(int nIterations, bool &bNoMore, vector<bool> &vbInliers, int &nInliers){ cv::Mat MLPnPsolver::iterate(int nIterations, bool &bNoMore, vector<bool> &vbInliers, int &nInliers){
bNoMore = false; bNoMore = false; // 已经达到最大迭代次数的标志
vbInliers.clear(); vbInliers.clear(); // 清除保存判断是否是内点的容器
nInliers=0; nInliers=0; // 当前次迭代时的内点数
// N为所有2D点的个数, mRansacMinInliers为正常退出RANSAC迭代过程中最少的inlier数
// Step 1: 判断如果2D点个数不足以启动RANSAC迭代过程的最小下限则退出
if(N<mRansacMinInliers) if(N<mRansacMinInliers)
{ {
bNoMore = true; bNoMore = true;
return cv::Mat(); return cv::Mat();
} }
// mvAllIndices为所有参与PnP的2D点的索引
// vAvailableIndices为每次从mvAllIndices中随机挑选mRansacMinSet组3D-2D对应点进行一次RANSAC
vector<size_t> vAvailableIndices; vector<size_t> vAvailableIndices;
// 当前的迭代次数id
int nCurrentIterations = 0; int nCurrentIterations = 0;
// Step 2: 正常迭代计算进行相机位姿估计,如果满足效果上限,直接返回最佳估计结果,否则就继续利用最小集(6个点)估计位姿
// 进行迭代的条件:
// 条件1: 历史进行的迭代次数少于最大迭代值
// 条件2: 当前进行的迭代次数少于当前函数给定的最大迭代值
while(mnIterations<mRansacMaxIts || nCurrentIterations<nIterations) while(mnIterations<mRansacMaxIts || nCurrentIterations<nIterations)
{ {
// 迭代次数更新加1直到达到最大迭代次数
nCurrentIterations++; nCurrentIterations++;
mnIterations++; mnIterations++;
// 清空已有的匹配点的计数,为新的一次迭代作准备
vAvailableIndices = mvAllIndices; vAvailableIndices = mvAllIndices;
//Bearing vectors and 3D points used for this ransac iteration //Bearing vectors and 3D points used for this ransac iteration
// 初始化单位向量和3D点给当前ransac迭代使用
bearingVectors_t bearingVecs(mRansacMinSet); bearingVectors_t bearingVecs(mRansacMinSet);
points_t p3DS(mRansacMinSet); points_t p3DS(mRansacMinSet);
vector<int> indexes(mRansacMinSet); vector<int> indexes(mRansacMinSet);
// Get min set of points // Get min set of points
// 选取最小集,从vAvailableIndices中选取mRansacMinSet个点进行操作这里应该是6
for(short i = 0; i < mRansacMinSet; ++i) for(short i = 0; i < mRansacMinSet; ++i)
{ {
// 在所有备选点中随机抽取一个通过随机抽取索引数组vAvailableIndices的索引[randi]来实现
int randi = DUtils::Random::RandomInt(0, vAvailableIndices.size()-1); int randi = DUtils::Random::RandomInt(0, vAvailableIndices.size()-1);
// vAvailableIndices[randi]才是备选点真正的索引值randi是索引数组的索引值不要搞混了
int idx = vAvailableIndices[randi]; int idx = vAvailableIndices[randi];
bearingVecs[i] = mvBearingVecs[idx]; bearingVecs[i] = mvBearingVecs[idx];
p3DS[i] = mvP3Dw[idx]; p3DS[i] = mvP3Dw[idx];
indexes[i] = i; indexes[i] = i;
// 把抽取出来的点从所有备选点数组里删除掉,概率论中不放回的操作
vAvailableIndices[randi] = vAvailableIndices.back(); vAvailableIndices[randi] = vAvailableIndices.back();
vAvailableIndices.pop_back(); vAvailableIndices.pop_back();
} } // 选取最小集
//By the moment, we are using MLPnP without covariance info //By the moment, we are using MLPnP without covariance info
// 目前为止还没有使用协方差的信息所以这里生成一个size=1的值为0的协方差矩阵
// |0 0 0|
// covs[0] = |0 0 0|
// |0 0 0|
// ? 为什么不用协方差的SVD分解计算耗时还是效果不明显
cov3_mats_t covs(1); cov3_mats_t covs(1);
//Result //Result
transformation_t result; transformation_t result;
// Compute camera pose // Compute camera pose
// 相机位姿估计MLPnP最主要的操作在这里
computePose(bearingVecs,p3DS,covs,indexes,result); computePose(bearingVecs,p3DS,covs,indexes,result);
//Save result //Save result
// 论文中12个待求值赋值保存在mRi中每个求解器都有保存各自的计算结果
mRi[0][0] = result(0,0); mRi[0][0] = result(0,0);
mRi[0][1] = result(0,1); mRi[0][1] = result(0,1);
mRi[0][2] = result(0,2); mRi[0][2] = result(0,2);
@ -162,15 +222,17 @@ namespace ORB_SLAM3 {
mti[0] = result(0,3);mti[1] = result(1,3);mti[2] = result(2,3); mti[0] = result(0,3);mti[1] = result(1,3);mti[2] = result(2,3);
// Check inliers // Check inliers
// 卡方检验内点和EPnP基本类似
CheckInliers(); CheckInliers();
if(mnInliersi>=mRansacMinInliers) if(mnInliersi>=mRansacMinInliers)
{ {
// If it is the best solution so far, save it // If it is the best solution so far, save it
// 如果该结果是目前内点数最多的,说明该结果是目前最好的,保存起来
if(mnInliersi>mnBestInliers) if(mnInliersi>mnBestInliers)
{ {
mvbBestInliers = mvbInliersi; mvbBestInliers = mvbInliersi; // 每个点是否是内点的标记
mnBestInliers = mnInliersi; mnBestInliers = mnInliersi; // 内点个数
cv::Mat Rcw(3,3,CV_64F,mRi); cv::Mat Rcw(3,3,CV_64F,mRi);
cv::Mat tcw(3,1,CV_64F,mti); cv::Mat tcw(3,1,CV_64F,mti);
@ -181,6 +243,7 @@ namespace ORB_SLAM3 {
tcw.copyTo(mBestTcw.rowRange(0,3).col(3)); tcw.copyTo(mBestTcw.rowRange(0,3).col(3));
} }
// 用新的内点对相机位姿精求解,提高位姿估计精度,这里如果有足够内点的话,函数直接返回该值,不再继续计算
if(Refine()) if(Refine())
{ {
nInliers = mnRefinedInliers; nInliers = mnRefinedInliers;
@ -194,8 +257,11 @@ namespace ORB_SLAM3 {
} }
} }
} } // 迭代
// Step 3: 选择最小集中效果最好的相机位姿估计结果,如果没有只能用6个点去计算这个值了
// 程序运行到这里说明Refine失败了说明精求解过程中内点的个数不满足最小阈值那就只能在当前结果中选择内点数最多的那个最小集
// 但是也意味着这样子的结果最终是用6个点来求出来的近似效果一般
if(mnIterations>=mRansacMaxIts) if(mnIterations>=mRansacMaxIts)
{ {
bNoMore=true; bNoMore=true;
@ -215,18 +281,30 @@ namespace ORB_SLAM3 {
return cv::Mat(); return cv::Mat();
} }
/**
* @brief RANSAC
*
* @param[in] probability 0.9
* @param[in] minInliers 8
* @param[in] maxIterations 300
* @param[in] minSet 6I > 5
* @param[in] epsilon epsilon0.4
* @param[in] th2
*
*/
void MLPnPsolver::SetRansacParameters(double probability, int minInliers, int maxIterations, int minSet, float epsilon, float th2){ void MLPnPsolver::SetRansacParameters(double probability, int minInliers, int maxIterations, int minSet, float epsilon, float th2){
mRansacProb = probability; mRansacProb = probability; // 模型最大概率值默认0.9
mRansacMinInliers = minInliers; mRansacMinInliers = minInliers; // 内点的最小阈值默认8
mRansacMaxIts = maxIterations; mRansacMaxIts = maxIterations; // 最大迭代次数默认300
mRansacEpsilon = epsilon; mRansacEpsilon = epsilon; // 理论最少内点个数这里是按照总数的比例计算所以epsilon是比例默认是0.4
mRansacMinSet = minSet; mRansacMinSet = minSet; // 每次采样六个点即最小集应该设置为6论文里面写着I > 5
N = mvP2D.size(); // number of correspondences N = mvP2D.size(); // number of correspondences
mvbInliersi.resize(N); mvbInliersi.resize(N); // 是否是内点的标记位
// Adjust Parameters according to number of correspondences // Adjust Parameters according to number of correspondences
// 计算最少个数点,选择(给定内点数, 最小集, 理论内点数)的最小值
int nMinInliers = N*mRansacEpsilon; int nMinInliers = N*mRansacEpsilon;
if(nMinInliers<mRansacMinInliers) if(nMinInliers<mRansacMinInliers)
nMinInliers=mRansacMinInliers; nMinInliers=mRansacMinInliers;
@ -234,10 +312,12 @@ namespace ORB_SLAM3 {
nMinInliers=minSet; nMinInliers=minSet;
mRansacMinInliers = nMinInliers; mRansacMinInliers = nMinInliers;
// 根据最终得到的"最小内点数"来调整 内点数/总体数 比例epsilon
if(mRansacEpsilon<(float)mRansacMinInliers/N) if(mRansacEpsilon<(float)mRansacMinInliers/N)
mRansacEpsilon=(float)mRansacMinInliers/N; mRansacEpsilon=(float)mRansacMinInliers/N;
// Set RANSAC iterations according to probability, epsilon, and max iterations // Set RANSAC iterations according to probability, epsilon, and max iterations
// 根据给出的各种参数计算RANSAC的理论迭代次数,并且敲定最终在迭代过程中使用的RANSAC最大迭代次数
int nIterations; int nIterations;
if(mRansacMinInliers==N) if(mRansacMinInliers==N)
@ -247,6 +327,7 @@ namespace ORB_SLAM3 {
mRansacMaxIts = max(1,min(nIterations,mRansacMaxIts)); mRansacMaxIts = max(1,min(nIterations,mRansacMaxIts));
// 计算不同图层上的特征点在进行内点检验的时候,所使用的不同判断误差阈值
mvMaxError.resize(mvSigma2.size()); mvMaxError.resize(mvSigma2.size());
for(size_t i=0; i<mvSigma2.size(); i++) for(size_t i=0; i<mvSigma2.size(); i++)
mvMaxError[i] = mvSigma2[i]*th2; mvMaxError[i] = mvSigma2[i]*th2;
@ -342,55 +423,124 @@ namespace ORB_SLAM3 {
} }
//MLPnP methods //MLPnP methods
/**
* @brief MLPnP姿
*
* @param[in] f
* @param[in] p 3D
* @param[in] covMats
* @param[in] indices
* @param[in] result 姿
*
*/
void MLPnPsolver::computePose(const bearingVectors_t &f, const points_t &p, const cov3_mats_t &covMats, void MLPnPsolver::computePose(const bearingVectors_t &f, const points_t &p, const cov3_mats_t &covMats,
const std::vector<int> &indices, transformation_t &result) { const std::vector<int> &indices, transformation_t &result) {
// Step 1: 判断点的数量是否满足计算条件,否则直接报错
// 因为每个观测值会产生2个残差所以至少需要6个点来计算公式12所以要检验当前的点个数是否满足大于5的条件
size_t numberCorrespondences = indices.size(); size_t numberCorrespondences = indices.size();
// 当numberCorrespondences不满足>5的条件时会发生错误(如果小于6根本进不来)
assert(numberCorrespondences > 5); assert(numberCorrespondences > 5);
// ? 用来标记是否满足平面条件平面情况下矩阵有相关性秩为2矩阵形式可以简化但需要跟多的约束求解
bool planar = false; bool planar = false;
// compute the nullspace of all vectors // compute the nullspace of all vectors
// compute the nullspace of all vectors
// step 2: 计算点的单位(方向向量)向量的零空间
// 利用公式7 Jvr(v) = null(v^T) = [r s]
// 给每个向量都开辟一个零空间,所以数量相等
std::vector<Eigen::MatrixXd> nullspaces(numberCorrespondences); std::vector<Eigen::MatrixXd> nullspaces(numberCorrespondences);
// 存储世界坐标系下空间点的矩阵3行N列N是numberCorrespondences即点的总个数
// |x1, x2, xn|
// points3 = |y1, y2, ..., yn|
// |z1, z2, zn|
Eigen::MatrixXd points3(3, numberCorrespondences); Eigen::MatrixXd points3(3, numberCorrespondences);
// 空间点向量
// |xi|
// points3v = |yi|
// |zi|
points_t points3v(numberCorrespondences); points_t points3v(numberCorrespondences);
// 单个空间点的齐次坐标矩阵TODO:没用到啊
// |xi|
// points4v = |yi|
// |zi|
// |1 |
points4_t points4v(numberCorrespondences); points4_t points4v(numberCorrespondences);
// numberCorrespondences不等于所有点而是提取出来的内点的数量其作为连续索引值对indices进行索引
// 因为内点的索引并非连续,想要方便遍历,必须用连续的索引值,
// 所以就用了indices[i]嵌套形式i表示内点数量numberCorrespondences范围内的连续形式
// indices里面保存的是不连续的内点的索引值
for (size_t i = 0; i < numberCorrespondences; i++) { for (size_t i = 0; i < numberCorrespondences; i++) {
// 当前空间点的单位向量,indices[i]是当前点坐标和向量的索引值,
bearingVector_t f_current = f[indices[i]]; bearingVector_t f_current = f[indices[i]];
// 取出当前点记录到 points3 空间点矩阵里
points3.col(i) = p[indices[i]]; points3.col(i) = p[indices[i]];
// nullspace of right vector // nullspace of right vector
// 求解方程 Jvr(v) = null(v^T) = [r s]
// A = U * S * V^T
// 这里只求解了V的完全解没有求解U
Eigen::JacobiSVD<Eigen::MatrixXd, Eigen::HouseholderQRPreconditioner> Eigen::JacobiSVD<Eigen::MatrixXd, Eigen::HouseholderQRPreconditioner>
svd_f(f_current.transpose(), Eigen::ComputeFullV); svd_f(f_current.transpose(), Eigen::ComputeFullV);
// 取特征值最小的那两个对应的2个特征向量
// |r1 s1|
// nullspaces = |r2 s2|
// |r3 s3|
nullspaces[i] = svd_f.matrixV().block(0, 1, 3, 2); nullspaces[i] = svd_f.matrixV().block(0, 1, 3, 2);
// 取出当前点记录到 points3v 空间点向量
points3v[i] = p[indices[i]]; points3v[i] = p[indices[i]];
} }
// Step 3: 通过计算S的秩来判断是在平面情况还是在正常情况
// 令S = M * M^T其中M = [p1,p2,...,pi],即 points3 空间点矩阵
////////////////////////////////////// //////////////////////////////////////
// 1. test if we have a planar scene // 1. test if we have a planar scene
// 在平面条件下会产生4个解因此需要另外判断和解决平面条件下的问题
////////////////////////////////////// //////////////////////////////////////
// 令S = M * M^T其中M = [p1,p2,...,pi],即 points3 空间点矩阵
// 如果矩阵S的秩为3且最小特征值不接近于0则不属于平面条件否则需要解决一下
Eigen::Matrix3d planarTest = points3 * points3.transpose(); Eigen::Matrix3d planarTest = points3 * points3.transpose();
Eigen::FullPivHouseholderQR<Eigen::Matrix3d> rankTest(planarTest); Eigen::FullPivHouseholderQR<Eigen::Matrix3d> rankTest(planarTest);
//int r, c; //int r, c;
//double minEigenVal = abs(eigen_solver.eigenvalues().real().minCoeff(&r, &c)); //double minEigenVal = abs(eigen_solver.eigenvalues().real().minCoeff(&r, &c));
// 特征旋转矩阵,用在平面条件下的计算
Eigen::Matrix3d eigenRot; Eigen::Matrix3d eigenRot;
eigenRot.setIdentity(); eigenRot.setIdentity();
// if yes -> transform points to new eigen frame // if yes -> transform points to new eigen frame
//if (minEigenVal < 1e-3 || minEigenVal == 0.0) //if (minEigenVal < 1e-3 || minEigenVal == 0.0)
//rankTest.setThreshold(1e-10); //rankTest.setThreshold(1e-10);
// 当矩阵S的秩为2时属于平面条件
if (rankTest.rank() == 2) { if (rankTest.rank() == 2) {
planar = true; planar = true;
// self adjoint is faster and more accurate than general eigen solvers // self adjoint is faster and more accurate than general eigen solvers
// also has closed form solution for 3x3 self-adjoint matrices // also has closed form solution for 3x3 self-adjoint matrices
// in addition this solver sorts the eigenvalues in increasing order // in addition this solver sorts the eigenvalues in increasing order
// 计算矩阵S的特征值和特征向量
Eigen::SelfAdjointEigenSolver<Eigen::Matrix3d> eigen_solver(planarTest); Eigen::SelfAdjointEigenSolver<Eigen::Matrix3d> eigen_solver(planarTest);
// 得到QR分解的结果
eigenRot = eigen_solver.eigenvectors().real(); eigenRot = eigen_solver.eigenvectors().real();
// 把eigenRot变成其转置矩阵即论文公式20的系数 R_S^T
eigenRot.transposeInPlace(); eigenRot.transposeInPlace();
// 公式20 pi' = R_S^T * pi
for (size_t i = 0; i < numberCorrespondences; i++) for (size_t i = 0; i < numberCorrespondences; i++)
points3.col(i) = eigenRot * points3.col(i); points3.col(i) = eigenRot * points3.col(i);
} }
////////////////////////////////////// //////////////////////////////////////
// 2. stochastic model // 2. stochastic model
////////////////////////////////////// //////////////////////////////////////
// Step 4: 计算随机模型中的协方差矩阵
// 但是作者并没有用到协方差信息
Eigen::SparseMatrix<double> P(2 * numberCorrespondences, Eigen::SparseMatrix<double> P(2 * numberCorrespondences,
2 * numberCorrespondences); 2 * numberCorrespondences);
bool use_cov = false; bool use_cov = false;
@ -398,6 +548,8 @@ namespace ORB_SLAM3 {
// if we do have covariance information // if we do have covariance information
// -> fill covariance matrix // -> fill covariance matrix
// 如果协方差矩阵的个数等于空间点的个数,说明前面已经计算好了,表示有协方差信息
// 目前版本是没有用到协方差信息的所以调用本函数前就把协方差矩阵个数置为1了
if (covMats.size() == numberCorrespondences) { if (covMats.size() == numberCorrespondences) {
use_cov = true; use_cov = true;
int l = 0; int l = 0;
@ -413,12 +565,22 @@ namespace ORB_SLAM3 {
} }
} }
// Step 5: 构造矩阵A来完成线性方程组的构建
////////////////////////////////////// //////////////////////////////////////
// 3. fill the design matrix A // 3. fill the design matrix A
////////////////////////////////////// //////////////////////////////////////
// 公式12设矩阵A则有 Au = 0
// u = [r11, r12, r13, r21, r22, r23, r31, r32, r33, t1, t2, t3]^T
// 对单位向量v的2个零空间向量做微分所以有[dr, ds]^T一个点有2行N个点就有2*N行
const int rowsA = 2 * numberCorrespondences; const int rowsA = 2 * numberCorrespondences;
// 对应上面u的列数因为旋转矩阵有9个元素加上平移矩阵3个元素总共12个元素
int colsA = 12; int colsA = 12;
Eigen::MatrixXd A; Eigen::MatrixXd A;
// 如果世界点位于分别跨2个坐标轴的平面上即所有世界点的一个元素是常数的时候可简单地忽略矩阵A中对应的列
// 而且这不影响问题的结构本身所以在计算公式20 pi' = R_S^T * pi的时候忽略了r11,r21,r31即第一列
// 对应的u只有9个元素 u = [r12, r13, r22, r23, r32, r33, t1, t2, t3]^T 所以A的列个数是9个
if (planar) { if (planar) {
colsA = 9; colsA = 9;
A = Eigen::MatrixXd(rowsA, 9); A = Eigen::MatrixXd(rowsA, 9);
@ -427,35 +589,37 @@ namespace ORB_SLAM3 {
A.setZero(); A.setZero();
// fill design matrix // fill design matrix
// 构造矩阵A分平面和非平面2种情况
if (planar) { if (planar) {
for (size_t i = 0; i < numberCorrespondences; ++i) { for (size_t i = 0; i < numberCorrespondences; ++i) {
// 列表示当前点的坐标
point_t pt3_current = points3.col(i); point_t pt3_current = points3.col(i);
// r12 // r12 r12 的系数 r1*py 和 s1*py
A(2 * i, 0) = nullspaces[i](0, 0) * pt3_current[1]; A(2 * i, 0) = nullspaces[i](0, 0) * pt3_current[1];
A(2 * i + 1, 0) = nullspaces[i](0, 1) * pt3_current[1]; A(2 * i + 1, 0) = nullspaces[i](0, 1) * pt3_current[1];
// r13 // r13 r13 的系数 r1*pz 和 s1*pz
A(2 * i, 1) = nullspaces[i](0, 0) * pt3_current[2]; A(2 * i, 1) = nullspaces[i](0, 0) * pt3_current[2];
A(2 * i + 1, 1) = nullspaces[i](0, 1) * pt3_current[2]; A(2 * i + 1, 1) = nullspaces[i](0, 1) * pt3_current[2];
// r22 // r22 r22 的系数 r2*py 和 s2*py
A(2 * i, 2) = nullspaces[i](1, 0) * pt3_current[1]; A(2 * i, 2) = nullspaces[i](1, 0) * pt3_current[1];
A(2 * i + 1, 2) = nullspaces[i](1, 1) * pt3_current[1]; A(2 * i + 1, 2) = nullspaces[i](1, 1) * pt3_current[1];
// r23 // r23 r23 的系数 r2*pz 和 s2*pz
A(2 * i, 3) = nullspaces[i](1, 0) * pt3_current[2]; A(2 * i, 3) = nullspaces[i](1, 0) * pt3_current[2];
A(2 * i + 1, 3) = nullspaces[i](1, 1) * pt3_current[2]; A(2 * i + 1, 3) = nullspaces[i](1, 1) * pt3_current[2];
// r32 // r32 r32 的系数 r3*py 和 s3*py
A(2 * i, 4) = nullspaces[i](2, 0) * pt3_current[1]; A(2 * i, 4) = nullspaces[i](2, 0) * pt3_current[1];
A(2 * i + 1, 4) = nullspaces[i](2, 1) * pt3_current[1]; A(2 * i + 1, 4) = nullspaces[i](2, 1) * pt3_current[1];
// r33 // r33 r33 的系数 r3*pz 和 s3*pz
A(2 * i, 5) = nullspaces[i](2, 0) * pt3_current[2]; A(2 * i, 5) = nullspaces[i](2, 0) * pt3_current[2];
A(2 * i + 1, 5) = nullspaces[i](2, 1) * pt3_current[2]; A(2 * i + 1, 5) = nullspaces[i](2, 1) * pt3_current[2];
// t1 // t1 t1 的系数 r1 和 s1
A(2 * i, 6) = nullspaces[i](0, 0); A(2 * i, 6) = nullspaces[i](0, 0);
A(2 * i + 1, 6) = nullspaces[i](0, 1); A(2 * i + 1, 6) = nullspaces[i](0, 1);
// t2 // t2 t2 的系数 r2 和 s2
A(2 * i, 7) = nullspaces[i](1, 0); A(2 * i, 7) = nullspaces[i](1, 0);
A(2 * i + 1, 7) = nullspaces[i](1, 1); A(2 * i + 1, 7) = nullspaces[i](1, 1);
// t3 // t3 t3 的系数 r3 和 s3
A(2 * i, 8) = nullspaces[i](2, 0); A(2 * i, 8) = nullspaces[i](2, 0);
A(2 * i + 1, 8) = nullspaces[i](2, 1); A(2 * i + 1, 8) = nullspaces[i](2, 1);
} }
@ -463,6 +627,7 @@ namespace ORB_SLAM3 {
for (size_t i = 0; i < numberCorrespondences; ++i) { for (size_t i = 0; i < numberCorrespondences; ++i) {
point_t pt3_current = points3.col(i); point_t pt3_current = points3.col(i);
// 不是平面的情况下,三个列向量都保留求解
// r11 // r11
A(2 * i, 0) = nullspaces[i](0, 0) * pt3_current[0]; A(2 * i, 0) = nullspaces[i](0, 0) * pt3_current[0];
A(2 * i + 1, 0) = nullspaces[i](0, 1) * pt3_current[0]; A(2 * i + 1, 0) = nullspaces[i](0, 1) * pt3_current[0];
@ -502,18 +667,26 @@ namespace ORB_SLAM3 {
} }
} }
// Step 6: 计算线性方程组的最小二乘解
////////////////////////////////////// //////////////////////////////////////
// 4. solve least squares // 4. solve least squares
////////////////////////////////////// //////////////////////////////////////
// 求解方程的最小二乘解
Eigen::MatrixXd AtPA; Eigen::MatrixXd AtPA;
if (use_cov) if (use_cov)
// 有协方差信息的情况下,一般方程是 A^T*P*A*u = N*u = 0
AtPA = A.transpose() * P * A; // setting up the full normal equations seems to be unstable AtPA = A.transpose() * P * A; // setting up the full normal equations seems to be unstable
else else
// 无协方差信息的情况下,一般方程是 A^T*A*u = N*u = 0
AtPA = A.transpose() * A; AtPA = A.transpose() * A;
// SVD分解满秩求解V
Eigen::JacobiSVD<Eigen::MatrixXd> svd_A(AtPA, Eigen::ComputeFullV); Eigen::JacobiSVD<Eigen::MatrixXd> svd_A(AtPA, Eigen::ComputeFullV);
// 解就是对应奇异值最小的列向量,即最后一列
Eigen::MatrixXd result1 = svd_A.matrixV().col(colsA - 1); Eigen::MatrixXd result1 = svd_A.matrixV().col(colsA - 1);
// Step 7: 根据平面和非平面情况下选择最终位姿解
//////////////////////////////// ////////////////////////////////
// now we treat the results differently, // now we treat the results differently,
// depending on the scene (planar or not) // depending on the scene (planar or not)
@ -526,30 +699,67 @@ namespace ORB_SLAM3 {
rotation_t tmp; rotation_t tmp;
// until now, we only estimated // until now, we only estimated
// row one and two of the transposed rotation matrix // row one and two of the transposed rotation matrix
// 暂时只估计了旋转矩阵的第1行和第2行先记录到tmp中
tmp << 0.0, result1(0, 0), result1(1, 0), tmp << 0.0, result1(0, 0), result1(1, 0),
0.0, result1(2, 0), result1(3, 0), 0.0, result1(2, 0), result1(3, 0),
0.0, result1(4, 0), result1(5, 0); 0.0, result1(4, 0), result1(5, 0);
//double scale = 1 / sqrt(tmp.col(1).norm() * tmp.col(2).norm()); //double scale = 1 / sqrt(tmp.col(1).norm() * tmp.col(2).norm());
// row 3 // row 3
// 第3行等于第1行和第2行的叉积这里应该是列后面转置后成了行
tmp.col(0) = tmp.col(1).cross(tmp.col(2)); tmp.col(0) = tmp.col(1).cross(tmp.col(2));
// 原来是:
// |r11 r12 r13|
// tmp = |r21 r22 r23|
// |r31 r32 r33|
// 转置变成:
// |r11 r21 r31|
// tmp = |r12 r22 r32|
// |r13 r23 r33|
tmp.transposeInPlace(); tmp.transposeInPlace();
double scale = 1.0 / std::sqrt(std::abs(tmp.col(1).norm() * tmp.col(2).norm())); double scale = 1.0 / std::sqrt(std::abs(tmp.col(1).norm() * tmp.col(2).norm()));
// find best rotation matrix in frobenius sense // find best rotation matrix in frobenius sense
// 利用Frobenious范数计算最佳的旋转矩阵利用公式(19) R = U_R*V_R^T
// 本质上采用矩阵将其元素平方将它们加在一起并对结果平方根。计算得出的数字是矩阵的Frobenious范数
// 由于列向量是单列矩阵行向量是单行矩阵所以这些矩阵的Frobenius范数等于向量的长度L
Eigen::JacobiSVD<Eigen::MatrixXd> svd_R_frob(tmp, Eigen::ComputeFullU | Eigen::ComputeFullV); Eigen::JacobiSVD<Eigen::MatrixXd> svd_R_frob(tmp, Eigen::ComputeFullU | Eigen::ComputeFullV);
rotation_t Rout1 = svd_R_frob.matrixU() * svd_R_frob.matrixV().transpose(); rotation_t Rout1 = svd_R_frob.matrixU() * svd_R_frob.matrixV().transpose();
// test if we found a good rotation matrix // test if we found a good rotation matrix
// 如果估计出来的旋转矩阵的行列式小于0则乘以-1EPnP也是同样的操作
if (Rout1.determinant() < 0) if (Rout1.determinant() < 0)
Rout1 *= -1.0; Rout1 *= -1.0;
// rotate this matrix back using the eigen frame // rotate this matrix back using the eigen frame
// ? 因为是在平面情况下计算的,估计出来的旋转矩阵是要做一个转换的,根据公式(21)R = Rs*R
// 其中Rs表示特征向量的旋转矩阵
// 注意eigenRot之前已经转置过了transposeInPlace()所以这里Rout1在之前也转置了即tmp.transposeInPlace()
Rout1 = eigenRot.transpose() * Rout1; Rout1 = eigenRot.transpose() * Rout1;
// 估计最终的平移矩阵,带尺度信息的,根据公式(17)t = t^ / three-party(||r1||*||r2||*||r3||)
// 这里是 t = t^ / sqrt(||r1||*||r2||)
translation_t t = scale * translation_t(result1(6, 0), result1(7, 0), result1(8, 0)); translation_t t = scale * translation_t(result1(6, 0), result1(7, 0), result1(8, 0));
// 把之前转置过来的矩阵再转回去,变成公式里面的形态:
// |r11 r12 r13|
// Rout1 = |r21 r22 r23|
// |r31 r32 r33|
Rout1.transposeInPlace(); Rout1.transposeInPlace();
// 这里乘以-1是为了计算4种结果
Rout1 *= -1; Rout1 *= -1;
if (Rout1.determinant() < 0.0) if (Rout1.determinant() < 0.0)
Rout1.col(2) *= -1; Rout1.col(2) *= -1;
// now we have to find the best out of 4 combinations // now we have to find the best out of 4 combinations
// |r11 r12 r13|
// R1 = |r21 r22 r23|
// |r31 r32 r33|
// |-r11 -r12 -r13|
// R2 = |-r21 -r22 -r23|
// |-r31 -r32 -r33|
rotation_t R1, R2; rotation_t R1, R2;
R1.col(0) = Rout1.col(0); R1.col(0) = Rout1.col(0);
R1.col(1) = Rout1.col(1); R1.col(1) = Rout1.col(1);
@ -558,6 +768,10 @@ namespace ORB_SLAM3 {
R2.col(1) = -Rout1.col(1); R2.col(1) = -Rout1.col(1);
R2.col(2) = Rout1.col(2); R2.col(2) = Rout1.col(2);
// |R1 t|
// Ts = |R1 -t|
// |R2 t|
// |R2 -t|
vector<transformation_t, Eigen::aligned_allocator<transformation_t>> Ts(4); vector<transformation_t, Eigen::aligned_allocator<transformation_t>> Ts(4);
Ts[0].block<3, 3>(0, 0) = R1; Ts[0].block<3, 3>(0, 0) = R1;
Ts[0].block<3, 1>(0, 3) = t; Ts[0].block<3, 1>(0, 3) = t;
@ -568,59 +782,124 @@ namespace ORB_SLAM3 {
Ts[3].block<3, 3>(0, 0) = R2; Ts[3].block<3, 3>(0, 0) = R2;
Ts[3].block<3, 1>(0, 3) = -t; Ts[3].block<3, 1>(0, 3) = -t;
// 遍历4种解
vector<double> normVal(4); vector<double> normVal(4);
for (int i = 0; i < 4; ++i) { for (int i = 0; i < 4; ++i) {
point_t reproPt; point_t reproPt;
double norms = 0.0; double norms = 0.0;
// 计算世界点p到切线空间v的投影的残差对应最小的就是最好的解
// 用前6个点来验证4种解的残差
for (int p = 0; p < 6; ++p) { for (int p = 0; p < 6; ++p) {
// 重投影的向量
reproPt = Ts[i].block<3, 3>(0, 0) * points3v[p] + Ts[i].block<3, 1>(0, 3); reproPt = Ts[i].block<3, 3>(0, 0) * points3v[p] + Ts[i].block<3, 1>(0, 3);
// 变成单位向量
reproPt = reproPt / reproPt.norm(); reproPt = reproPt / reproPt.norm();
// f[indices[p]] 是当前空间点的单位向量
// 利用欧氏距离来表示重投影向量(观测)和当前空间点向量(实际)的偏差
// 即两个n维向量a(x11,x12,…,x1n)与 b(x21,x22,…,x2n)间的欧氏距离
norms += (1.0 - reproPt.transpose() * f[indices[p]]); norms += (1.0 - reproPt.transpose() * f[indices[p]]);
} }
// 统计每种解的误差和第i个解的误差和放入对应的变量normVal[i]
normVal[i] = norms; normVal[i] = norms;
} }
// 搜索容器中的最小值,并返回该值对应的指针
std::vector<double>::iterator std::vector<double>::iterator
findMinRepro = std::min_element(std::begin(normVal), std::end(normVal)); findMinRepro = std::min_element(std::begin(normVal), std::end(normVal));
// 计算容器头指针到最小值指针的距离,即可作为该最小值的索引值
int idx = std::distance(std::begin(normVal), findMinRepro); int idx = std::distance(std::begin(normVal), findMinRepro);
// 得到最终相机位姿估计的结果
Rout = Ts[idx].block<3, 3>(0, 0); Rout = Ts[idx].block<3, 3>(0, 0);
tout = Ts[idx].block<3, 1>(0, 3); tout = Ts[idx].block<3, 1>(0, 3);
} else // non-planar } else // non-planar
{ {
rotation_t tmp; rotation_t tmp;
// 从AtPA的SVD分解中得到旋转矩阵先存下来
// 注意这里的顺序是和公式16不同的
// |r11 r21 r31|
// tmp = |r12 r22 r32|
// |r13 r23 r33|
tmp << result1(0, 0), result1(3, 0), result1(6, 0), tmp << result1(0, 0), result1(3, 0), result1(6, 0),
result1(1, 0), result1(4, 0), result1(7, 0), result1(1, 0), result1(4, 0), result1(7, 0),
result1(2, 0), result1(5, 0), result1(8, 0); result1(2, 0), result1(5, 0), result1(8, 0);
// get the scale // get the scale
// 计算尺度,根据公式(17)t = t^ / three-party(||r1||*||r2||*||r3||)
double scale = 1.0 / double scale = 1.0 /
std::pow(std::abs(tmp.col(0).norm() * tmp.col(1).norm() * tmp.col(2).norm()), 1.0 / 3.0); std::pow(std::abs(tmp.col(0).norm() * tmp.col(1).norm() * tmp.col(2).norm()), 1.0 / 3.0);
//double scale = 1.0 / std::sqrt(std::abs(tmp.col(0).norm() * tmp.col(1).norm())); //double scale = 1.0 / std::sqrt(std::abs(tmp.col(0).norm() * tmp.col(1).norm()));
// find best rotation matrix in frobenius sense // find best rotation matrix in frobenius sense
// 利用Frobenious范数计算最佳的旋转矩阵利用公式(19) R = U_R*V_R^T
Eigen::JacobiSVD<Eigen::MatrixXd> svd_R_frob(tmp, Eigen::ComputeFullU | Eigen::ComputeFullV); Eigen::JacobiSVD<Eigen::MatrixXd> svd_R_frob(tmp, Eigen::ComputeFullU | Eigen::ComputeFullV);
Rout = svd_R_frob.matrixU() * svd_R_frob.matrixV().transpose(); Rout = svd_R_frob.matrixU() * svd_R_frob.matrixV().transpose();
// test if we found a good rotation matrix // test if we found a good rotation matrix
// 如果估计出来的旋转矩阵的行列式小于0则乘以-1
if (Rout.determinant() < 0) if (Rout.determinant() < 0)
Rout *= -1.0; Rout *= -1.0;
// scale translation // scale translation
// 从相机坐标系到世界坐标系的转换关系是 lambda*v = R*pi+t
// 从世界坐标系到相机坐标系的转换关系是 pi = R^T*v-R^Tt
// 旋转矩阵的性质 R^-1 = R^T
// 所以在下面的计算中需要计算从世界坐标系到相机坐标系的转换这里tout = -R^T*t下面再计算前半部分R^T*v
// 先恢复平移部分的尺度再计算
tout = Rout * (scale * translation_t(result1(9, 0), result1(10, 0), result1(11, 0))); tout = Rout * (scale * translation_t(result1(9, 0), result1(10, 0), result1(11, 0)));
// find correct direction in terms of reprojection error, just take the first 6 correspondences // find correct direction in terms of reprojection error, just take the first 6 correspondences
// 非平面情况下一共有2种解R,t和R,-t
// 利用前6个点计算重投影误差选择残差最小的一个解
vector<double> error(2); vector<double> error(2);
vector<Eigen::Matrix4d, Eigen::aligned_allocator<Eigen::Matrix4d>> Ts(2); vector<Eigen::Matrix4d, Eigen::aligned_allocator<Eigen::Matrix4d>> Ts(2);
for (int s = 0; s < 2; ++s) { for (int s = 0; s < 2; ++s) {
// 初始化error的值为0
error[s] = 0.0; error[s] = 0.0;
// |1 0 0 0|
// Ts[s] = |0 1 0 0|
// |0 0 1 0|
// |0 0 0 1|
Ts[s] = Eigen::Matrix4d::Identity(); Ts[s] = Eigen::Matrix4d::Identity();
// |. . . 0|
// Ts[s] = |. Rout . 0|
// |. . . 0|
// |0 0 0 1|
Ts[s].block<3, 3>(0, 0) = Rout; Ts[s].block<3, 3>(0, 0) = Rout;
if (s == 0) if (s == 0)
// |. . . . |
// Ts[s] = |. Rout . tout|
// |. . . . |
// |0 0 0 1 |
Ts[s].block<3, 1>(0, 3) = tout; Ts[s].block<3, 1>(0, 3) = tout;
else else
// |. . . . |
// Ts[s] = |. Rout . -tout|
// |. . . . |
// |0 0 0 1 |
Ts[s].block<3, 1>(0, 3) = -tout; Ts[s].block<3, 1>(0, 3) = -tout;
// 为了避免Eigen中aliasing的问题后面在计算矩阵的逆的时候需要添加eval()条件
// a = a.transpose(); //error: aliasing
// a = a.transpose().eval(); //ok
// a.transposeInPlace(); //ok
// Eigen中aliasing指的是在赋值表达式的左右两边存在矩阵的重叠区域这种情况下有可能得到非预期的结果。
// 如mat = 2*mat或者mat = mat.transpose()第一个例子中的alias是没有问题的而第二的例子则会导致非预期的计算结果。
Ts[s] = Ts[s].inverse().eval(); Ts[s] = Ts[s].inverse().eval();
for (int p = 0; p < 6; ++p) { for (int p = 0; p < 6; ++p) {
// 从世界坐标系到相机坐标系的转换关系是 pi = R^T*v-R^Tt
// Ts[s].block<3, 3>(0, 0) * points3v[p] = Rout = R^T*v
// Ts[s].block<3, 1>(0, 3) = tout = -R^Tt
bearingVector_t v = Ts[s].block<3, 3>(0, 0) * points3v[p] + Ts[s].block<3, 1>(0, 3); bearingVector_t v = Ts[s].block<3, 3>(0, 0) * points3v[p] + Ts[s].block<3, 1>(0, 3);
// 变成单位向量
v = v / v.norm(); v = v / v.norm();
// 计算重投影向量(观测)和当前空间点向量(实际)的偏差
error[s] += (1.0 - v.transpose() * f[indices[p]]); error[s] += (1.0 - v.transpose() * f[indices[p]]);
} }
} }
// 选择残差最小的解作为最终解
if (error[0] < error[1]) if (error[0] < error[1])
tout = Ts[0].block<3, 1>(0, 3); tout = Ts[0].block<3, 1>(0, 3);
else else
@ -629,10 +908,18 @@ namespace ORB_SLAM3 {
} }
// Step 8: 利用高斯牛顿法对位姿进行精确求解,提高位姿解的精度
////////////////////////////////////// //////////////////////////////////////
// 5. gauss newton // 5. gauss newton
////////////////////////////////////// //////////////////////////////////////
// 求解非线性方程之前,需要得到罗德里格斯参数,来表达李群(SO3) -> 李代数(so3)的对数映射
rodrigues_t omega = rot2rodrigues(Rout); rodrigues_t omega = rot2rodrigues(Rout);
// |r1|
// |r2|
// minx = |r3|
// |t1|
// |t2|
// |t3|
Eigen::VectorXd minx(6); Eigen::VectorXd minx(6);
minx[0] = omega[0]; minx[0] = omega[0];
minx[1] = omega[1]; minx[1] = omega[1];
@ -641,11 +928,15 @@ namespace ORB_SLAM3 {
minx[4] = tout[1]; minx[4] = tout[1];
minx[5] = tout[2]; minx[5] = tout[2];
// 利用高斯牛顿迭代法来提炼相机位姿 pose
mlpnp_gn(minx, points3v, nullspaces, P, use_cov); mlpnp_gn(minx, points3v, nullspaces, P, use_cov);
// 最终输出的结果
Rout = rodrigues2rot(rodrigues_t(minx[0], minx[1], minx[2])); Rout = rodrigues2rot(rodrigues_t(minx[0], minx[1], minx[2]));
tout = translation_t(minx[3], minx[4], minx[5]); tout = translation_t(minx[3], minx[4], minx[5]);
// result inverse as opengv uses this convention // result inverse as opengv uses this convention
// 这里是用来计算世界坐标系到相机坐标系的转换所以是Pc=R^T*Pw-R^T*t反变换
result.block<3, 3>(0, 0) = Rout;//Rout.transpose(); result.block<3, 3>(0, 0) = Rout;//Rout.transpose();
result.block<3, 1>(0, 3) = tout;//-result.block<3, 3>(0, 0) * tout; result.block<3, 1>(0, 3) = tout;//-result.block<3, 3>(0, 0) * tout;
} }
@ -684,11 +975,25 @@ namespace ORB_SLAM3 {
return omega; return omega;
} }
/**
* @brief MLPnP
* @param[in] x
* @param[in] pts 3D
* @param[in] nullspaces
* @param[in] Kll
* @param[in] use_cov 使
*/
void MLPnPsolver::mlpnp_gn(Eigen::VectorXd &x, const points_t &pts, const std::vector<Eigen::MatrixXd> &nullspaces, void MLPnPsolver::mlpnp_gn(Eigen::VectorXd &x, const points_t &pts, const std::vector<Eigen::MatrixXd> &nullspaces,
const Eigen::SparseMatrix<double> Kll, bool use_cov) { const Eigen::SparseMatrix<double> Kll, bool use_cov) {
// 计算观测值数量
const int numObservations = pts.size(); const int numObservations = pts.size();
// 未知量是旋转向量和平移向量即R和t总共6个未知参数
const int numUnknowns = 6; const int numUnknowns = 6;
// check redundancy // check redundancy
// 检查观测数量是否满足计算条件因为每个观测值都提供了2个约束即r和s所以这里乘以2
assert((2 * numObservations - numUnknowns) > 0); assert((2 * numObservations - numUnknowns) > 0);
// ============= // =============
@ -751,6 +1056,15 @@ namespace ORB_SLAM3 {
// result // result
} }
/**
* @brief Jacobian
* @param[in] x
* @param[in] pts 3D
* @param[in] nullspaces
* @param[in] r f(x)
* @param[out] fjac f(x)Jacobian
* @param[in] getJacs Jacobian
*/
void MLPnPsolver::mlpnp_residuals_and_jacs(const Eigen::VectorXd &x, const points_t &pts, void MLPnPsolver::mlpnp_residuals_and_jacs(const Eigen::VectorXd &x, const points_t &pts,
const std::vector<Eigen::MatrixXd> &nullspaces, Eigen::VectorXd &r, const std::vector<Eigen::MatrixXd> &nullspaces, Eigen::VectorXd &r,
Eigen::MatrixXd &fjac, bool getJacs) { Eigen::MatrixXd &fjac, bool getJacs) {
@ -800,6 +1114,15 @@ namespace ORB_SLAM3 {
} }
} }
/**
* @brief Jacobian
* @param[in] pt 3D
* @param[in] nullspace_r r
* @param[in] nullspace_r s
* @param[in] w w
* @param[in] t t
* @param[out] jacs Jacobian
*/
void MLPnPsolver::mlpnpJacs(const point_t& pt, const Eigen::Vector3d& nullspace_r, void MLPnPsolver::mlpnpJacs(const point_t& pt, const Eigen::Vector3d& nullspace_r,
const Eigen::Vector3d& nullspace_s, const rodrigues_t& w, const Eigen::Vector3d& nullspace_s, const rodrigues_t& w,
const translation_t& t, Eigen::MatrixXd& jacs){ const translation_t& t, Eigen::MatrixXd& jacs){

View File

@ -3430,12 +3430,15 @@ bool Tracking::Relocalization()
{ {
Verbose::PrintMess("Starting relocalization", Verbose::VERBOSITY_NORMAL); Verbose::PrintMess("Starting relocalization", Verbose::VERBOSITY_NORMAL);
// Compute Bag of Words Vector // Compute Bag of Words Vector
// Step 1: 计算当前帧特征点的Bow映射
mCurrentFrame.ComputeBoW(); mCurrentFrame.ComputeBoW();
// Relocalization is performed when tracking is lost // Relocalization is performed when tracking is lost
// Track Lost: Query KeyFrame Database for keyframe candidates for relocalisation // Track Lost: Query KeyFrame Database for keyframe candidates for relocalisation
// Step 2找到与当前帧相似的候选关键帧组
vector<KeyFrame*> vpCandidateKFs = mpKeyFrameDB->DetectRelocalizationCandidates(&mCurrentFrame, mpAtlas->GetCurrentMap()); vector<KeyFrame*> vpCandidateKFs = mpKeyFrameDB->DetectRelocalizationCandidates(&mCurrentFrame, mpAtlas->GetCurrentMap());
// 如果没有候选关键帧,则退出
if(vpCandidateKFs.empty()) { if(vpCandidateKFs.empty()) {
Verbose::PrintMess("There are not candidates", Verbose::VERBOSITY_NORMAL); Verbose::PrintMess("There are not candidates", Verbose::VERBOSITY_NORMAL);
return false; return false;
@ -3447,17 +3450,22 @@ bool Tracking::Relocalization()
// If enough matches are found we setup a PnP solver // If enough matches are found we setup a PnP solver
ORBmatcher matcher(0.75,true); ORBmatcher matcher(0.75,true);
// 每个关键帧的解算器
vector<MLPnPsolver*> vpMLPnPsolvers; vector<MLPnPsolver*> vpMLPnPsolvers;
vpMLPnPsolvers.resize(nKFs); vpMLPnPsolvers.resize(nKFs);
// 每个关键帧和当前帧中特征点的匹配关系
vector<vector<MapPoint*> > vvpMapPointMatches; vector<vector<MapPoint*> > vvpMapPointMatches;
vvpMapPointMatches.resize(nKFs); vvpMapPointMatches.resize(nKFs);
// 放弃某个关键帧的标记
vector<bool> vbDiscarded; vector<bool> vbDiscarded;
vbDiscarded.resize(nKFs); vbDiscarded.resize(nKFs);
// 有效的候选关键帧数目
int nCandidates=0; int nCandidates=0;
// Step 3遍历所有的候选关键帧通过BoW进行快速匹配用匹配结果初始化PnP Solver
for(int i=0; i<nKFs; i++) for(int i=0; i<nKFs; i++)
{ {
KeyFrame* pKF = vpCandidateKFs[i]; KeyFrame* pKF = vpCandidateKFs[i];
@ -3465,6 +3473,7 @@ bool Tracking::Relocalization()
vbDiscarded[i] = true; vbDiscarded[i] = true;
else else
{ {
// 当前帧和候选关键帧用BoW进行快速匹配匹配结果记录在vvpMapPointMatchesnmatches表示匹配的数目
int nmatches = matcher.SearchByBoW(pKF,mCurrentFrame,vvpMapPointMatches[i]); int nmatches = matcher.SearchByBoW(pKF,mCurrentFrame,vvpMapPointMatches[i]);
if(nmatches<15) if(nmatches<15)
{ {
@ -3473,8 +3482,18 @@ bool Tracking::Relocalization()
} }
else else
{ {
// 如果匹配数目够用用匹配结果初始化MLPnPsolver
// ? 为什么用MLPnP? 因为考虑了鱼眼相机模型,解耦某些关系?
// 参考论文《MLPNP-A REAL-TIME MAXIMUM LIKELIHOOD SOLUTION TO THE PERSPECTIVE-N-POINT PROBLEM》
MLPnPsolver* pSolver = new MLPnPsolver(mCurrentFrame,vvpMapPointMatches[i]); MLPnPsolver* pSolver = new MLPnPsolver(mCurrentFrame,vvpMapPointMatches[i]);
pSolver->SetRansacParameters(0.99,10,300,6,0.5,5.991); //This solver needs at least 6 points // 构造函数调用了一遍,这里重新设置参数
pSolver->SetRansacParameters(
0.99, // 模型最大概率值默认0.9
10, // 内点的最小阈值默认8
300, // 最大迭代次数默认300
6, // 最小集每次采样六个点即最小集应该设置为6论文里面写着I > 5
0.5, // 理论最少内点个数这里是按照总数的比例计算所以epsilon是比例默认是0.4
5.991); // 卡方检验阈值 //This solver needs at least 6 points
vpMLPnPsolvers[i] = pSolver; vpMLPnPsolvers[i] = pSolver;
} }
} }
@ -3482,9 +3501,13 @@ bool Tracking::Relocalization()
// Alternatively perform some iterations of P4P RANSAC // Alternatively perform some iterations of P4P RANSAC
// Until we found a camera pose supported by enough inliers // Until we found a camera pose supported by enough inliers
// 足够的内点才能匹配使用PNP算法MLPnP需要至少6个点
// 是否已经找到相匹配的关键帧的标志
bool bMatch = false; bool bMatch = false;
ORBmatcher matcher2(0.9,true); ORBmatcher matcher2(0.9,true);
// Step 4: 通过一系列操作,直到找到能够匹配上的关键帧
// 为什么搞这么复杂?答:是担心误闭环
while(nCandidates>0 && !bMatch) while(nCandidates>0 && !bMatch)
{ {
for(int i=0; i<nKFs; i++) for(int i=0; i<nKFs; i++)
@ -3493,14 +3516,22 @@ bool Tracking::Relocalization()
continue; continue;
// Perform 5 Ransac Iterations // Perform 5 Ransac Iterations
// 内点标记
vector<bool> vbInliers; vector<bool> vbInliers;
// 内点数
int nInliers; int nInliers;
// 表示RANSAC已经没有更多的迭代次数可用 -- 也就是说数据不够好RANSAC也已经尽力了。。。
bool bNoMore; bool bNoMore;
// Step 4.1通过MLPnP算法估计姿态迭代5次
MLPnPsolver* pSolver = vpMLPnPsolvers[i]; MLPnPsolver* pSolver = vpMLPnPsolvers[i];
// PnP算法的入口函数
cv::Mat Tcw = pSolver->iterate(5,bNoMore,vbInliers,nInliers); cv::Mat Tcw = pSolver->iterate(5,bNoMore,vbInliers,nInliers);
// If Ransac reachs max. iterations discard keyframe // If Ransac reachs max. iterations discard keyframe
// bNoMore 为true 表示已经超过了RANSAC最大迭代次数就放弃当前关键帧
if(bNoMore) if(bNoMore)
{ {
vbDiscarded[i]=true; vbDiscarded[i]=true;
@ -3510,12 +3541,15 @@ bool Tracking::Relocalization()
// If a Camera Pose is computed, optimize // If a Camera Pose is computed, optimize
if(!Tcw.empty()) if(!Tcw.empty())
{ {
// Step 4.2如果MLPnP 计算出了位姿对内点进行BA优化
Tcw.copyTo(mCurrentFrame.mTcw); Tcw.copyTo(mCurrentFrame.mTcw);
// MLPnP 里RANSAC后的内点的集合
set<MapPoint*> sFound; set<MapPoint*> sFound;
const int np = vbInliers.size(); const int np = vbInliers.size();
// 遍历所有内点
for(int j=0; j<np; j++) for(int j=0; j<np; j++)
{ {
if(vbInliers[j]) if(vbInliers[j])
@ -3527,43 +3561,66 @@ bool Tracking::Relocalization()
mCurrentFrame.mvpMapPoints[j]=NULL; mCurrentFrame.mvpMapPoints[j]=NULL;
} }
// 只优化位姿,不优化地图点的坐标,返回的是内点的数量
int nGood = Optimizer::PoseOptimization(&mCurrentFrame); int nGood = Optimizer::PoseOptimization(&mCurrentFrame);
// 如果优化之后的内点数目不多,跳过了当前候选关键帧,但是却没有放弃当前帧的重定位
if(nGood<10) if(nGood<10)
continue; continue;
// 删除外点对应的地图点,这里直接设为空指针
for(int io =0; io<mCurrentFrame.N; io++) for(int io =0; io<mCurrentFrame.N; io++)
if(mCurrentFrame.mvbOutlier[io]) if(mCurrentFrame.mvbOutlier[io])
mCurrentFrame.mvpMapPoints[io]=static_cast<MapPoint*>(NULL); mCurrentFrame.mvpMapPoints[io]=static_cast<MapPoint*>(NULL);
// If few inliers, search by projection in a coarse window and optimize again // If few inliers, search by projection in a coarse window and optimize again
// Step 4.3:如果内点较少,则通过投影的方式对之前未匹配的点进行匹配,再进行优化求解
// 前面的匹配关系是用词袋匹配过程得到的
if(nGood<50) if(nGood<50)
{ {
int nadditional =matcher2.SearchByProjection(mCurrentFrame,vpCandidateKFs[i],sFound,10,100); // 通过投影的方式将关键帧中未匹配的地图点投影到当前帧中, 生成新的匹配
int nadditional =matcher2.SearchByProjection(
mCurrentFrame, // 当前帧
vpCandidateKFs[i], // 关键帧
sFound, // 已经找到的地图点集合不会用于PNP
10, // 窗口阈值,会乘以金字塔尺度
100); // 匹配的ORB描述子距离应该小于这个阈值
if(nadditional+nGood>=50) if(nadditional+nGood>=50)
{ {
// 根据投影匹配的结果再次采用3D-2D pnp BA优化位姿
nGood = Optimizer::PoseOptimization(&mCurrentFrame); nGood = Optimizer::PoseOptimization(&mCurrentFrame);
// If many inliers but still not enough, search by projection again in a narrower window // If many inliers but still not enough, search by projection again in a narrower window
// the camera has been already optimized with many points // the camera has been already optimized with many points
// Step 4.4如果BA后内点数还是比较少(<50)但是还不至于太少(>30),可以挽救一下, 最后垂死挣扎
// 重新执行上一步 4.3的过程,只不过使用更小的搜索窗口
// 这里的位姿已经使用了更多的点进行了优化,应该更准,所以使用更小的窗口搜索
if(nGood>30 && nGood<50) if(nGood>30 && nGood<50)
{ {
// 用更小窗口、更严格的描述子阈值,重新进行投影搜索匹配
sFound.clear(); sFound.clear();
for(int ip =0; ip<mCurrentFrame.N; ip++) for(int ip =0; ip<mCurrentFrame.N; ip++)
if(mCurrentFrame.mvpMapPoints[ip]) if(mCurrentFrame.mvpMapPoints[ip])
sFound.insert(mCurrentFrame.mvpMapPoints[ip]); sFound.insert(mCurrentFrame.mvpMapPoints[ip]);
nadditional =matcher2.SearchByProjection(mCurrentFrame,vpCandidateKFs[i],sFound,3,64); nadditional =matcher2.SearchByProjection(
mCurrentFrame, // 当前帧
vpCandidateKFs[i], // 候选的关键帧
sFound, // 已经找到的地图点不会用于PNP
3, // 新的窗口阈值,会乘以金字塔尺度
64); // 匹配的ORB描述子距离应该小于这个阈值
// Final optimization // Final optimization
// 如果成功挽救回来匹配数目达到要求最后BA优化一下
if(nGood+nadditional>=50) if(nGood+nadditional>=50)
{ {
nGood = Optimizer::PoseOptimization(&mCurrentFrame); nGood = Optimizer::PoseOptimization(&mCurrentFrame);
//更新地图点
for(int io =0; io<mCurrentFrame.N; io++) for(int io =0; io<mCurrentFrame.N; io++)
if(mCurrentFrame.mvbOutlier[io]) if(mCurrentFrame.mvbOutlier[io])
mCurrentFrame.mvpMapPoints[io]=NULL; mCurrentFrame.mvpMapPoints[io]=NULL;
} }
//如果还是不能够满足就放弃了
} }
} }
} }
@ -3573,6 +3630,7 @@ bool Tracking::Relocalization()
if(nGood>=50) if(nGood>=50)
{ {
bMatch = true; bMatch = true;
// 只要有一个候选关键帧重定位成功,就退出循环,不考虑其他候选关键帧了
break; break;
} }
} }