提交 0a63ab36 编写于 作者: T Tong Ke 提交者: Alexander Alekhin

Merge pull request #8301 from tonyke1993:p3p_alg

New p3p algorithm (accepted by CVPR 2017) (#8301)

* add p3p source code

* indent 4

* update publication info

* fix filename

* interface done

* plug in done, test needed

* debugging

* for test

* a working version

* clean p3p code

* test

* test

* fix warning, blank line

* apply patch from @catree

* add reference info

* namespace, indent 4

* static solveQuartic

* put small functions to anonymous namespace
上级 2561c596
......@@ -897,3 +897,10 @@
year={2010},
publisher={Springer}
}
@INPROCEEDINGS{Ke17,
author = {Ke, Tong and Roumeliotis, Stergios},
title = {An Efficient Algebraic Solution to the Perspective-Three-Point Problem},
booktitle = {Computer Vision and Pattern Recognition (CVPR), 2017 IEEE Conference on},
year = {2017},
organization = {IEEE}
}
......@@ -236,8 +236,8 @@ enum { SOLVEPNP_ITERATIVE = 0,
SOLVEPNP_EPNP = 1, //!< EPnP: Efficient Perspective-n-Point Camera Pose Estimation @cite lepetit2009epnp
SOLVEPNP_P3P = 2, //!< Complete Solution Classification for the Perspective-Three-Point Problem @cite gao2003complete
SOLVEPNP_DLS = 3, //!< A Direct Least-Squares (DLS) Method for PnP @cite hesch2011direct
SOLVEPNP_UPNP = 4 //!< Exhaustive Linearization for Robust Camera Pose and Focal Length Estimation @cite penate2013exhaustive
SOLVEPNP_UPNP = 4, //!< Exhaustive Linearization for Robust Camera Pose and Focal Length Estimation @cite penate2013exhaustive
SOLVEPNP_AP3P = 5 //!< An Efficient Algebraic Solution to the Perspective-Three-Point Problem @cite Ke17
};
enum { CALIB_CB_ADAPTIVE_THRESH = 1,
......
#include "ap3p.h"
#include <cmath>
#include <complex>
using namespace std;
namespace {
void solveQuartic(const double *factors, double *realRoots) {
const double &a4 = factors[0];
const double &a3 = factors[1];
const double &a2 = factors[2];
const double &a1 = factors[3];
const double &a0 = factors[4];
double a4_2 = a4 * a4;
double a3_2 = a3 * a3;
double a4_3 = a4_2 * a4;
double a2a4 = a2 * a4;
double p4 = (8 * a2a4 - 3 * a3_2) / (8 * a4_2);
double q4 = (a3_2 * a3 - 4 * a2a4 * a3 + 8 * a1 * a4_2) / (8 * a4_3);
double r4 = (256 * a0 * a4_3 - 3 * (a3_2 * a3_2) - 64 * a1 * a3 * a4_2 + 16 * a2a4 * a3_2) / (256 * (a4_3 * a4));
double p3 = ((p4 * p4) / 12 + r4) / 3; // /=-3
double q3 = (72 * r4 * p4 - 2 * p4 * p4 * p4 - 27 * q4 * q4) / 432; // /=2
double t; // *=2
complex<double> w;
if (q3 >= 0)
w = -sqrt(static_cast<complex<double> >(q3 * q3 - p3 * p3 * p3)) - q3;
else
w = sqrt(static_cast<complex<double> >(q3 * q3 - p3 * p3 * p3)) - q3;
if (w.imag() == 0.0) {
w.real(cbrt(w.real()));
t = 2.0 * (w.real() + p3 / w.real());
} else {
w = pow(w, 1.0 / 3);
t = 4.0 * w.real();
}
complex<double> sqrt_2m = sqrt(static_cast<complex<double> >(-2 * p4 / 3 + t));
double B_4A = -a3 / (4 * a4);
double complex1 = 4 * p4 / 3 + t;
complex<double> complex2 = 2 * q4 / sqrt_2m;
double sqrt_2m_rh = sqrt_2m.real() / 2;
double sqrt1 = sqrt(-(complex1 + complex2)).real() / 2;
realRoots[0] = B_4A + sqrt_2m_rh + sqrt1;
realRoots[1] = B_4A + sqrt_2m_rh - sqrt1;
double sqrt2 = sqrt(-(complex1 - complex2)).real() / 2;
realRoots[2] = B_4A - sqrt_2m_rh + sqrt2;
realRoots[3] = B_4A - sqrt_2m_rh - sqrt2;
}
void polishQuarticRoots(const double *coeffs, double *roots) {
const int iterations = 2;
for (int i = 0; i < iterations; ++i) {
for (int j = 0; j < 4; ++j) {
double error =
(((coeffs[0] * roots[j] + coeffs[1]) * roots[j] + coeffs[2]) * roots[j] + coeffs[3]) * roots[j] +
coeffs[4];
double
derivative =
((4 * coeffs[0] * roots[j] + 3 * coeffs[1]) * roots[j] + 2 * coeffs[2]) * roots[j] + coeffs[3];
roots[j] -= error / derivative;
}
}
}
inline void vect_cross(const double *a, const double *b, double *result) {
result[0] = a[1] * b[2] - a[2] * b[1];
result[1] = -(a[0] * b[2] - a[2] * b[0]);
result[2] = a[0] * b[1] - a[1] * b[0];
}
inline double vect_dot(const double *a, const double *b) {
return a[0] * b[0] + a[1] * b[1] + a[2] * b[2];
}
inline double vect_norm(const double *a) {
return sqrt(a[0] * a[0] + a[1] * a[1] + a[2] * a[2]);
}
inline void vect_scale(const double s, const double *a, double *result) {
result[0] = a[0] * s;
result[1] = a[1] * s;
result[2] = a[2] * s;
}
inline void vect_sub(const double *a, const double *b, double *result) {
result[0] = a[0] - b[0];
result[1] = a[1] - b[1];
result[2] = a[2] - b[2];
}
inline void vect_divide(const double *a, const double d, double *result) {
result[0] = a[0] / d;
result[1] = a[1] / d;
result[2] = a[2] / d;
}
inline void mat_mult(const double a[3][3], const double b[3][3], double result[3][3]) {
result[0][0] = a[0][0] * b[0][0] + a[0][1] * b[1][0] + a[0][2] * b[2][0];
result[0][1] = a[0][0] * b[0][1] + a[0][1] * b[1][1] + a[0][2] * b[2][1];
result[0][2] = a[0][0] * b[0][2] + a[0][1] * b[1][2] + a[0][2] * b[2][2];
result[1][0] = a[1][0] * b[0][0] + a[1][1] * b[1][0] + a[1][2] * b[2][0];
result[1][1] = a[1][0] * b[0][1] + a[1][1] * b[1][1] + a[1][2] * b[2][1];
result[1][2] = a[1][0] * b[0][2] + a[1][1] * b[1][2] + a[1][2] * b[2][2];
result[2][0] = a[2][0] * b[0][0] + a[2][1] * b[1][0] + a[2][2] * b[2][0];
result[2][1] = a[2][0] * b[0][1] + a[2][1] * b[1][1] + a[2][2] * b[2][1];
result[2][2] = a[2][0] * b[0][2] + a[2][1] * b[1][2] + a[2][2] * b[2][2];
}
}
namespace cv {
void ap3p::init_inverse_parameters() {
inv_fx = 1. / fx;
inv_fy = 1. / fy;
cx_fx = cx / fx;
cy_fy = cy / fy;
}
ap3p::ap3p(cv::Mat cameraMatrix) {
if (cameraMatrix.depth() == CV_32F)
init_camera_parameters<float>(cameraMatrix);
else
init_camera_parameters<double>(cameraMatrix);
init_inverse_parameters();
}
ap3p::ap3p(double _fx, double _fy, double _cx, double _cy) {
fx = _fx;
fy = _fy;
cx = _cx;
cy = _cy;
init_inverse_parameters();
}
// This algorithm is from "Tong Ke, Stergios Roumeliotis, An Efficient Algebraic Solution to the Perspective-Three-Point Problem" (Accepted by CVPR 2017)
// See https://arxiv.org/pdf/1701.08237.pdf
// featureVectors: The 3 bearing measurements (normalized) stored as column vectors
// worldPoints: The positions of the 3 feature points stored as column vectors
// solutionsR: 4 possible solutions of rotation matrix of the world w.r.t the camera frame
// solutionsT: 4 possible solutions of translation of the world origin w.r.t the camera frame
int ap3p::computePoses(const double featureVectors[3][3],
const double worldPoints[3][3],
double solutionsR[4][3][3],
double solutionsT[4][3]) {
//world point vectors
double w1[3] = {worldPoints[0][0], worldPoints[1][0], worldPoints[2][0]};
double w2[3] = {worldPoints[0][1], worldPoints[1][1], worldPoints[2][1]};
double w3[3] = {worldPoints[0][2], worldPoints[1][2], worldPoints[2][2]};
// k1
double u0[3];
vect_sub(w1, w2, u0);
double nu0 = vect_norm(u0);
double k1[3];
vect_divide(u0, nu0, k1);
// bi
double b1[3] = {featureVectors[0][0], featureVectors[1][0], featureVectors[2][0]};
double b2[3] = {featureVectors[0][1], featureVectors[1][1], featureVectors[2][1]};
double b3[3] = {featureVectors[0][2], featureVectors[1][2], featureVectors[2][2]};
// k3,tz
double k3[3];
vect_cross(b1, b2, k3);
double nk3 = vect_norm(k3);
vect_divide(k3, nk3, k3);
double tz[3];
vect_cross(b1, k3, tz);
// ui,vi
double v1[3];
vect_cross(b1, b3, v1);
double v2[3];
vect_cross(b2, b3, v2);
double u1[3];
vect_sub(w1, w3, u1);
// coefficients related terms
double u1k1 = vect_dot(u1, k1);
double k3b3 = vect_dot(k3, b3);
// f1i
double f11 = k3b3;
double f13 = vect_dot(k3, v1);
double f15 = -u1k1 * f11;
//delta
double nl[3];
vect_cross(u1, k1, nl);
double delta = vect_norm(nl);
vect_divide(nl, delta, nl);
f11 *= delta;
f13 *= delta;
// f2i
double u2k1 = u1k1 - nu0;
double f21 = vect_dot(tz, v2);
double f22 = nk3 * k3b3;
double f23 = vect_dot(k3, v2);
double f24 = u2k1 * f22;
double f25 = -u2k1 * f21;
f21 *= delta;
f22 *= delta;
f23 *= delta;
double g1 = f13 * f22;
double g2 = f13 * f25 - f15 * f23;
double g3 = f11 * f23 - f13 * f21;
double g4 = -f13 * f24;
double g5 = f11 * f22;
double g6 = f11 * f25 - f15 * f21;
double g7 = -f15 * f24;
double coeffs[5] = {g5 * g5 + g1 * g1 + g3 * g3,
2 * (g5 * g6 + g1 * g2 + g3 * g4),
g6 * g6 + 2 * g5 * g7 + g2 * g2 + g4 * g4 - g1 * g1 - g3 * g3,
2 * (g6 * g7 - g1 * g2 - g3 * g4),
g7 * g7 - g2 * g2 - g4 * g4};
double s[4];
solveQuartic(coeffs, s);
polishQuarticRoots(coeffs, s);
double temp[3];
vect_cross(k1, nl, temp);
double Ck1nl[3][3] =
{{k1[0], nl[0], temp[0]},
{k1[1], nl[1], temp[1]},
{k1[2], nl[2], temp[2]}};
double Cb1k3tzT[3][3] =
{{b1[0], b1[1], b1[2]},
{k3[0], k3[1], k3[2]},
{tz[0], tz[1], tz[2]}};
double b3p[3];
vect_scale((delta / k3b3), b3, b3p);
int nb_solutions = 0;
for (int i = 0; i < 4; ++i) {
double ctheta1p = s[i];
if (abs(ctheta1p) > 1)
continue;
double stheta1p = sqrt(1 - ctheta1p * ctheta1p);
stheta1p = (k3b3 > 0) ? stheta1p : -stheta1p;
double ctheta3 = g1 * ctheta1p + g2;
double stheta3 = g3 * ctheta1p + g4;
double ntheta3 = stheta1p / ((g5 * ctheta1p + g6) * ctheta1p + g7);
ctheta3 *= ntheta3;
stheta3 *= ntheta3;
double C13[3][3] =
{{ctheta3, 0, -stheta3},
{stheta1p * stheta3, ctheta1p, stheta1p * ctheta3},
{ctheta1p * stheta3, -stheta1p, ctheta1p * ctheta3}};
double temp_matrix[3][3];
double R[3][3];
mat_mult(Ck1nl, C13, temp_matrix);
mat_mult(temp_matrix, Cb1k3tzT, R);
// R' * p3
double rp3[3] =
{w3[0] * R[0][0] + w3[1] * R[1][0] + w3[2] * R[2][0],
w3[0] * R[0][1] + w3[1] * R[1][1] + w3[2] * R[2][1],
w3[0] * R[0][2] + w3[1] * R[1][2] + w3[2] * R[2][2]};
double pxstheta1p[3];
vect_scale(stheta1p, b3p, pxstheta1p);
vect_sub(pxstheta1p, rp3, solutionsT[nb_solutions]);
solutionsR[nb_solutions][0][0] = R[0][0];
solutionsR[nb_solutions][1][0] = R[0][1];
solutionsR[nb_solutions][2][0] = R[0][2];
solutionsR[nb_solutions][0][1] = R[1][0];
solutionsR[nb_solutions][1][1] = R[1][1];
solutionsR[nb_solutions][2][1] = R[1][2];
solutionsR[nb_solutions][0][2] = R[2][0];
solutionsR[nb_solutions][1][2] = R[2][1];
solutionsR[nb_solutions][2][2] = R[2][2];
nb_solutions++;
}
return nb_solutions;
}
bool ap3p::solve(cv::Mat &R, cv::Mat &tvec, const cv::Mat &opoints, const cv::Mat &ipoints) {
CV_INSTRUMENT_REGION()
double rotation_matrix[3][3], translation[3];
std::vector<double> points;
if (opoints.depth() == ipoints.depth()) {
if (opoints.depth() == CV_32F)
extract_points<cv::Point3f, cv::Point2f>(opoints, ipoints, points);
else
extract_points<cv::Point3d, cv::Point2d>(opoints, ipoints, points);
} else if (opoints.depth() == CV_32F)
extract_points<cv::Point3f, cv::Point2d>(opoints, ipoints, points);
else
extract_points<cv::Point3d, cv::Point2f>(opoints, ipoints, points);
bool result = solve(rotation_matrix, translation, points[0], points[1], points[2], points[3], points[4], points[5],
points[6], points[7], points[8], points[9], points[10], points[11], points[12], points[13],
points[14],
points[15], points[16], points[17], points[18], points[19]);
cv::Mat(3, 1, CV_64F, translation).copyTo(tvec);
cv::Mat(3, 3, CV_64F, rotation_matrix).copyTo(R);
return result;
}
bool
ap3p::solve(double R[3][3], double t[3], double mu0, double mv0, double X0, double Y0, double Z0, double mu1,
double mv1,
double X1, double Y1, double Z1, double mu2, double mv2, double X2, double Y2, double Z2, double mu3,
double mv3, double X3, double Y3, double Z3) {
double Rs[4][3][3], ts[4][3];
int n = solve(Rs, ts, mu0, mv0, X0, Y0, Z0, mu1, mv1, X1, Y1, Z1, mu2, mv2, X2, Y2, Z2);
if (n == 0)
return false;
int ns = 0;
double min_reproj = 0;
for (int i = 0; i < n; i++) {
double X3p = Rs[i][0][0] * X3 + Rs[i][0][1] * Y3 + Rs[i][0][2] * Z3 + ts[i][0];
double Y3p = Rs[i][1][0] * X3 + Rs[i][1][1] * Y3 + Rs[i][1][2] * Z3 + ts[i][1];
double Z3p = Rs[i][2][0] * X3 + Rs[i][2][1] * Y3 + Rs[i][2][2] * Z3 + ts[i][2];
double mu3p = cx + fx * X3p / Z3p;
double mv3p = cy + fy * Y3p / Z3p;
double reproj = (mu3p - mu3) * (mu3p - mu3) + (mv3p - mv3) * (mv3p - mv3);
if (i == 0 || min_reproj > reproj) {
ns = i;
min_reproj = reproj;
}
}
for (int i = 0; i < 3; i++) {
for (int j = 0; j < 3; j++)
R[i][j] = Rs[ns][i][j];
t[i] = ts[ns][i];
}
return true;
}
int ap3p::solve(double R[4][3][3], double t[4][3], double mu0, double mv0, double X0, double Y0, double Z0, double mu1,
double mv1, double X1, double Y1, double Z1, double mu2, double mv2, double X2, double Y2, double Z2) {
double mk0, mk1, mk2;
double norm;
mu0 = inv_fx * mu0 - cx_fx;
mv0 = inv_fy * mv0 - cy_fy;
norm = sqrt(mu0 * mu0 + mv0 * mv0 + 1);
mk0 = 1. / norm;
mu0 *= mk0;
mv0 *= mk0;
mu1 = inv_fx * mu1 - cx_fx;
mv1 = inv_fy * mv1 - cy_fy;
norm = sqrt(mu1 * mu1 + mv1 * mv1 + 1);
mk1 = 1. / norm;
mu1 *= mk1;
mv1 *= mk1;
mu2 = inv_fx * mu2 - cx_fx;
mv2 = inv_fy * mv2 - cy_fy;
norm = sqrt(mu2 * mu2 + mv2 * mv2 + 1);
mk2 = 1. / norm;
mu2 *= mk2;
mv2 *= mk2;
double featureVectors[3][3] = {{mu0, mu1, mu2},
{mv0, mv1, mv2},
{mk0, mk1, mk2}};
double worldPoints[3][3] = {{X0, X1, X2},
{Y0, Y1, Y2},
{Z0, Z1, Z2}};
return computePoses(featureVectors, worldPoints, R, t);
}
}
\ No newline at end of file
#ifndef P3P_P3P_H
#define P3P_P3P_H
#include "precomp.hpp"
namespace cv {
class ap3p {
private:
template<typename T>
void init_camera_parameters(const cv::Mat &cameraMatrix) {
cx = cameraMatrix.at<T>(0, 2);
cy = cameraMatrix.at<T>(1, 2);
fx = cameraMatrix.at<T>(0, 0);
fy = cameraMatrix.at<T>(1, 1);
}
template<typename OpointType, typename IpointType>
void extract_points(const cv::Mat &opoints, const cv::Mat &ipoints, std::vector<double> &points) {
points.clear();
points.resize(20);
for (int i = 0; i < 4; i++) {
points[i * 5] = ipoints.at<IpointType>(i).x * fx + cx;
points[i * 5 + 1] = ipoints.at<IpointType>(i).y * fy + cy;
points[i * 5 + 2] = opoints.at<OpointType>(i).x;
points[i * 5 + 3] = opoints.at<OpointType>(i).y;
points[i * 5 + 4] = opoints.at<OpointType>(i).z;
}
}
void init_inverse_parameters();
double fx, fy, cx, cy;
double inv_fx, inv_fy, cx_fx, cy_fy;
public:
ap3p() {}
ap3p(double fx, double fy, double cx, double cy);
ap3p(cv::Mat cameraMatrix);
bool solve(cv::Mat &R, cv::Mat &tvec, const cv::Mat &opoints, const cv::Mat &ipoints);
int solve(double R[4][3][3], double t[4][3],
double mu0, double mv0, double X0, double Y0, double Z0,
double mu1, double mv1, double X1, double Y1, double Z1,
double mu2, double mv2, double X2, double Y2, double Z2);
bool solve(double R[3][3], double t[3],
double mu0, double mv0, double X0, double Y0, double Z0,
double mu1, double mv1, double X1, double Y1, double Z1,
double mu2, double mv2, double X2, double Y2, double Z2,
double mu3, double mv3, double X3, double Y3, double Z3);
// This algorithm is from "Tong Ke, Stergios Roumeliotis, An Efficient Algebraic Solution to the Perspective-Three-Point Problem" (Accepted by CVPR 2017)
// See https://arxiv.org/pdf/1701.08237.pdf
// featureVectors: 3 bearing measurements (normalized) stored as column vectors
// worldPoints: Positions of the 3 feature points stored as column vectors
// solutionsR: 4 possible solutions of rotation matrix of the world w.r.t the camera frame
// solutionsT: 4 possible solutions of translation of the world origin w.r.t the camera frame
int computePoses(const double featureVectors[3][3], const double worldPoints[3][3], double solutionsR[4][3][3],
double solutionsT[4][3]);
};
}
#endif //P3P_P3P_H
......@@ -45,6 +45,7 @@
#include "dls.h"
#include "epnp.h"
#include "p3p.h"
#include "ap3p.h"
#include "opencv2/calib3d/calib3d_c.h"
#include <iostream>
......@@ -118,6 +119,18 @@ bool solvePnP( InputArray _opoints, InputArray _ipoints,
if (result)
Rodrigues(R, rvec);
}
else if (flags == SOLVEPNP_AP3P)
{
CV_Assert( npoints == 4);
Mat undistortedPoints;
undistortPoints(ipoints, undistortedPoints, cameraMatrix, distCoeffs);
ap3p P3Psolver(cameraMatrix);
Mat R;
result = P3Psolver.solve(R, tvec, opoints, undistortedPoints);
if (result)
Rodrigues(R, rvec);
}
else if (flags == SOLVEPNP_ITERATIVE)
{
CvMat c_objectPoints = opoints, c_imagePoints = ipoints;
......@@ -291,7 +304,8 @@ bool solvePnPRansac(InputArray _opoints, InputArray _ipoints,
opoints_inliers.resize(npoints1);
ipoints_inliers.resize(npoints1);
result = solvePnP(opoints_inliers, ipoints_inliers, cameraMatrix,
distCoeffs, rvec, tvec, false, flags == SOLVEPNP_P3P ? SOLVEPNP_EPNP : flags) ? 1 : -1;
distCoeffs, rvec, tvec, false,
(flags == SOLVEPNP_P3P || flags == SOLVEPNP_AP3P) ? SOLVEPNP_EPNP : flags) ? 1 : -1;
}
if( result <= 0 || _local_model.rows <= 0)
......
......@@ -57,6 +57,7 @@ public:
eps[SOLVEPNP_ITERATIVE] = 1.0e-2;
eps[SOLVEPNP_EPNP] = 1.0e-2;
eps[SOLVEPNP_P3P] = 1.0e-2;
eps[SOLVEPNP_AP3P] = 1.0e-2;
eps[SOLVEPNP_DLS] = 1.0e-2;
eps[SOLVEPNP_UPNP] = 1.0e-2;
totalTestsCount = 10;
......@@ -161,7 +162,7 @@ protected:
points.resize(pointsCount);
generate3DPointCloud(points);
const int methodsCount = 5;
const int methodsCount = 6;
RNG rng = ts->get_rng();
......@@ -189,7 +190,7 @@ protected:
}
}
}
double eps[5];
double eps[6];
int totalTestsCount;
};
......@@ -201,6 +202,7 @@ public:
eps[SOLVEPNP_ITERATIVE] = 1.0e-6;
eps[SOLVEPNP_EPNP] = 1.0e-6;
eps[SOLVEPNP_P3P] = 1.0e-4;
eps[SOLVEPNP_AP3P] = 1.0e-4;
eps[SOLVEPNP_DLS] = 1.0e-4;
eps[SOLVEPNP_UPNP] = 1.0e-4;
totalTestsCount = 1000;
......@@ -222,7 +224,7 @@ protected:
generatePose(trueRvec, trueTvec, rng);
std::vector<Point3f> opoints;
if (method == 2)
if (method == 2 || method == 5)
{
opoints = std::vector<Point3f>(points.begin(), points.begin()+4);
}
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册