CPose3DQuatPDF.cpp

Go to the documentation of this file.

/* +------------------------------------------------------------------------+
   |                     Mobile Robot Programming Toolkit (MRPT)            |
   |                          https://www.mrpt.org/                         |
   |                                                                        |
   | Copyright (c) 2005-2019, Individual contributors, see AUTHORS file     |
   | See: https://www.mrpt.org/Authors - All rights reserved.               |
   | Released under BSD License. See: https://www.mrpt.org/License          |
   +------------------------------------------------------------------------+ */

#include "poses-precomp.h"  // Precompiled headers

#include <mrpt/poses/CPose3DPDFGaussian.h>
#include <mrpt/poses/CPose3DQuatPDF.h>
#include <mrpt/poses/CPose3DQuatPDFGaussian.h>
#include <mrpt/serialization/CArchive.h>
#include <Eigen/Dense>

using namespace mrpt::poses;
using namespace mrpt::math;
using namespace std;

IMPLEMENTS_VIRTUAL_SERIALIZABLE(CPose3DQuatPDF, CSerializable, mrpt::poses)

/*---------------------------------------------------------------
                    copyFrom2D
  ---------------------------------------------------------------*/
CPose3DQuatPDF* CPose3DQuatPDF::createFrom2D(const CPosePDF& o)
{
    MRPT_START

    CPose3DPDFGaussian q;
    q.copyFrom(o);

    return new CPose3DQuatPDFGaussian(q);

    MRPT_END
}

/*---------------------------------------------------------------
                    jacobiansPoseComposition
 ---------------------------------------------------------------*/
void CPose3DQuatPDF::jacobiansPoseComposition(
    const CPose3DQuat& x, const CPose3DQuat& u, CMatrixDouble77& df_dx,
    CMatrixDouble77& df_du, CPose3DQuat* out_x_oplus_u)
{
    // For the derivation of the formulas, see the tech. report cited in the
    // header file.
    const double qr = x.quat().r();
    const double qx = x.quat().x();
    const double qx2 = square(qx);
    const double qy = x.quat().y();
    const double qy2 = square(qy);
    const double qz = x.quat().z();
    const double qz2 = square(qz);

    const double ax = u.x();
    const double ay = u.y();
    const double az = u.z();
    const double q2r = u.quat().r();
    const double q2x = u.quat().x();
    const double q2y = u.quat().y();
    const double q2z = u.quat().z();

    CPose3DQuat x_plus_u = x + u;  // needed for the normalization Jacobian:
    CMatrixDouble44 norm_jacob(UNINITIALIZED_MATRIX);
    x_plus_u.quat().normalizationJacobian(norm_jacob);

    CMatrixDouble44 norm_jacob_x(UNINITIALIZED_MATRIX);
    x.quat().normalizationJacobian(norm_jacob_x);

    // df_dx ===================================================
    df_dx.setZero();

    // first part 3x7:  df_{qr} / dp
    df_dx(0, 0) = 1;
    df_dx(1, 1) = 1;
    df_dx(2, 2) = 1;

    alignas(MRPT_MAX_STATIC_ALIGN_BYTES)
        const double vals2[3 * 4] = {2 * (-qz * ay + qy * az),
                                     2 * (qy * ay + qz * az),
                                     2 * (-2 * qy * ax + qx * ay + qr * az),
                                     2 * (-2 * qz * ax - qr * ay + qx * az),

                                     2 * (qz * ax - qx * az),
                                     2 * (qy * ax - 2 * qx * ay - qr * az),
                                     2 * (qx * ax + qz * az),
                                     2 * (qr * ax - 2 * qz * ay + qy * az),

                                     2 * (-qy * ax + qx * ay),
                                     2 * (qz * ax + qr * ay - 2 * qx * az),
                                     2 * (-qr * ax + qz * ay - 2 * qy * az),
                                     2 * (qx * ax + qy * ay)};

    // df_dx(0:3,3:7) = vals2 * NORM_JACOB
    df_dx.block<3, 4>(0, 3).noalias() =
        (CMatrixFixed<double, 3, 4>(vals2) * norm_jacob_x).eval();

    // second part:
    {
        alignas(MRPT_MAX_STATIC_ALIGN_BYTES) const double aux44_data[4 * 4] = {
            q2r, -q2x, -q2y, -q2z, q2x, q2r, q2z,  -q2y,
            q2y, -q2z, q2r,  q2x,  q2z, q2y, -q2x, q2r};

        df_dx.block<4, 4>(3, 3).noalias() =
            (norm_jacob * CMatrixFixed<double, 4, 4>(aux44_data)).asEigen();
    }

    // df_du ===================================================
    df_du.setZero();

    // first part 3x3:  df_{qr} / da
    df_du(0, 0) = 1 - 2 * (qy2 + qz2);
    df_du(0, 1) = 2 * (qx * qy - qr * qz);
    df_du(0, 2) = 2 * (qr * qy + qx * qz);

    df_du(1, 0) = 2 * (qr * qz + qx * qy);
    df_du(1, 1) = 1 - 2 * (qx2 + qz2);
    df_du(1, 2) = 2 * (qy * qz - qr * qx);

    df_du(2, 0) = 2 * (qx * qz - qr * qy);
    df_du(2, 1) = 2 * (qr * qx + qy * qz);
    df_du(2, 2) = 1 - 2 * (qx2 + qy2);

    // Second part:
    {
        alignas(MRPT_MAX_STATIC_ALIGN_BYTES) const double aux44_data[4 * 4] = {
            qr, -qx, -qy, -qz, qx, qr,  -qz, qy,
            qy, qz,  qr,  -qx, qz, -qy, qx,  qr};

        df_du.block<4, 4>(3, 3).noalias() =
            (norm_jacob * CMatrixFixed<double, 4, 4>(aux44_data)).asEigen();
    }

    if (out_x_oplus_u) *out_x_oplus_u = x_plus_u;
}