CRangeBearingKFSLAM.cpp

Go to the documentation of this file.

/* +------------------------------------------------------------------------+
   |                     Mobile Robot Programming Toolkit (MRPT)            |
   |                          https://www.mrpt.org/                         |
   |                                                                        |
   | Copyright (c) 2005-2020, 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 "slam-precomp.h"  // Precompiled headers

// ----------------------------------------------------------------------------------------
// For the theory behind this implementation, see the technical report in:
//
//            https://www.mrpt.org/6D-SLAM
// ----------------------------------------------------------------------------------------

#include <mrpt/math/ops_containers.h>
#include <mrpt/math/utils.h>
#include <mrpt/math/wrap2pi.h>
#include <mrpt/obs/CActionRobotMovement3D.h>
#include <mrpt/opengl/CEllipsoid3D.h>
#include <mrpt/opengl/CSetOfObjects.h>
#include <mrpt/opengl/stock_objects.h>
#include <mrpt/poses/CPose3DQuatPDFGaussian.h>
#include <mrpt/poses/CPosePDF.h>
#include <mrpt/poses/CPosePDFGaussian.h>
#include <mrpt/slam/CRangeBearingKFSLAM.h>
#include <mrpt/system/CTicTac.h>
#include <mrpt/system/os.h>
#include <Eigen/Dense>

using namespace mrpt::slam;
using namespace mrpt::obs;
using namespace mrpt::maps;
using namespace mrpt::math;
using namespace mrpt::poses;
using namespace mrpt::system;
using namespace mrpt;
using namespace std;

/*---------------------------------------------------------------
                            Constructor
  ---------------------------------------------------------------*/
CRangeBearingKFSLAM::CRangeBearingKFSLAM()
    : options(),
      m_action(),
      m_SF(),
      m_IDs(),
      mapPartitioner(),
      m_SFs(),
      m_lastPartitionSet()
{
    reset();
}

/*---------------------------------------------------------------
                            reset
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::reset()
{
    m_action.reset();
    m_SF.reset();
    m_IDs.clear();
    m_SFs.clear();
    mapPartitioner.clear();
    m_lastPartitionSet.clear();

    // -----------------------
    // INIT KF STATE
    m_xkk.assign(get_vehicle_size(), 0);  // State: 6D pose (x,y,z)=(0,0,0)
    m_xkk[3] = 1.0;  // (qr,qx,qy,qz)=(1,0,0,0)

    // Initial cov:  nullptr diagonal -> perfect knowledge.
    m_pkk.setSize(get_vehicle_size(), get_vehicle_size());
    m_pkk.setZero();
    // -----------------------

    // Use SF-based matching (faster & easier for bearing-range observations
    // with ID).
    mapPartitioner.options.simil_method = smOBSERVATION_OVERLAP;
}

/*---------------------------------------------------------------
                            Destructor
  ---------------------------------------------------------------*/
CRangeBearingKFSLAM::~CRangeBearingKFSLAM() = default;
/*---------------------------------------------------------------
                            getCurrentRobotPose
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::getCurrentRobotPose(
    CPose3DQuatPDFGaussian& out_robotPose) const
{
    MRPT_START

    ASSERT_(m_xkk.size() >= 7);

    // Copy xyz+quat: (explicitly unroll the loop)
    out_robotPose.mean.m_coords[0] = m_xkk[0];
    out_robotPose.mean.m_coords[1] = m_xkk[1];
    out_robotPose.mean.m_coords[2] = m_xkk[2];
    out_robotPose.mean.m_quat[0] = m_xkk[3];
    out_robotPose.mean.m_quat[1] = m_xkk[4];
    out_robotPose.mean.m_quat[2] = m_xkk[5];
    out_robotPose.mean.m_quat[3] = m_xkk[6];

    // and cov:
    out_robotPose.cov = m_pkk.blockCopy<7, 7>(0, 0);

    MRPT_END
}

/*---------------------------------------------------------------
                            getCurrentRobotPoseMean
  ---------------------------------------------------------------*/
mrpt::poses::CPose3DQuat CRangeBearingKFSLAM::getCurrentRobotPoseMean() const
{
    CPose3DQuat q(mrpt::math::UNINITIALIZED_QUATERNION);
    ASSERTDEB_(m_xkk.size() >= 7);
    // Copy xyz+quat: (explicitly unroll the loop)
    q.m_coords[0] = m_xkk[0];
    q.m_coords[1] = m_xkk[1];
    q.m_coords[2] = m_xkk[2];
    q.m_quat[0] = m_xkk[3];
    q.m_quat[1] = m_xkk[4];
    q.m_quat[2] = m_xkk[5];
    q.m_quat[3] = m_xkk[6];

    return q;
}

/*---------------------------------------------------------------
                            getCurrentState
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::getCurrentState(
    CPose3DQuatPDFGaussian& out_robotPose,
    std::vector<TPoint3D>& out_landmarksPositions,
    std::map<unsigned int, CLandmark::TLandmarkID>& out_landmarkIDs,
    CVectorDouble& out_fullState, CMatrixDouble& out_fullCovariance) const
{
    MRPT_START

    ASSERT_(size_t(m_xkk.size()) >= get_vehicle_size());

    // Copy xyz+quat: (explicitly unroll the loop)
    out_robotPose.mean.m_coords[0] = m_xkk[0];
    out_robotPose.mean.m_coords[1] = m_xkk[1];
    out_robotPose.mean.m_coords[2] = m_xkk[2];
    out_robotPose.mean.m_quat[0] = m_xkk[3];
    out_robotPose.mean.m_quat[1] = m_xkk[4];
    out_robotPose.mean.m_quat[2] = m_xkk[5];
    out_robotPose.mean.m_quat[3] = m_xkk[6];

    // and cov:
    out_robotPose.cov = m_pkk.blockCopy<7, 7>(0, 0);

    // Landmarks:
    ASSERT_(((m_xkk.size() - get_vehicle_size()) % get_feature_size()) == 0);
    size_t i, nLMs = (m_xkk.size() - get_vehicle_size()) / get_feature_size();
    out_landmarksPositions.resize(nLMs);
    for (i = 0; i < nLMs; i++)
    {
        out_landmarksPositions[i].x =
            m_xkk[get_vehicle_size() + i * get_feature_size() + 0];
        out_landmarksPositions[i].y =
            m_xkk[get_vehicle_size() + i * get_feature_size() + 1];
        out_landmarksPositions[i].z =
            m_xkk[get_vehicle_size() + i * get_feature_size() + 2];
    }  // end for i

    // IDs:
    out_landmarkIDs = m_IDs.getInverseMap();  // m_IDs_inverse;

    // Full state:
    out_fullState.resize(m_xkk.size());
    std::copy(m_xkk.begin(), m_xkk.end(), out_fullState.begin());
    // Full cov:
    out_fullCovariance = m_pkk;

    MRPT_END
}

/*---------------------------------------------------------------
                        processActionObservation
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::processActionObservation(
    CActionCollection::Ptr& action, CSensoryFrame::Ptr& SF)
{
    MRPT_START

    m_action = action;
    m_SF = SF;

    // Sanity check:
    ASSERT_(
        m_IDs.size() ==
        (m_pkk.cols() - get_vehicle_size()) / get_feature_size());

    // ===================================================================================================================
    // Here's the meat!: Call the main method for the KF algorithm, which will
    // call all the callback methods as required:
    // ===================================================================================================================
    runOneKalmanIteration();

    // =============================================================
    //  ADD TO SFs SEQUENCE
    // =============================================================
    CPose3DQuatPDFGaussian q(UNINITIALIZED_QUATERNION);
    this->getCurrentRobotPose(q);
    CPose3DPDFGaussian::Ptr auxPosePDF = CPose3DPDFGaussian::Create(q);

    if (options.create_simplemap)
    {
        m_SFs.insert(CPose3DPDF::Ptr(auxPosePDF), SF);
    }

    // =============================================================
    //  UPDATE THE PARTITION GRAPH EXPERIMENT
    // =============================================================
    if (options.doPartitioningExperiment)
    {
        if (options.partitioningMethod == 0)
        {
            // Use spectral-graph technique:
            mapPartitioner.addMapFrame(*SF, *auxPosePDF);

            vector<std::vector<uint32_t>> partitions;
            mapPartitioner.updatePartitions(partitions);
            m_lastPartitionSet = partitions;
        }
        else
        {
            // Fixed partitions every K observations:
            vector<std::vector<uint32_t>> partitions;
            std::vector<uint32_t> tmpCluster;

            ASSERT_(options.partitioningMethod > 1);
            size_t N = options.partitioningMethod;

            for (size_t i = 0; i < m_SFs.size(); i++)
            {
                tmpCluster.push_back(i);
                if ((i % N) == 0)
                {
                    partitions.push_back(tmpCluster);
                    tmpCluster.clear();
                    tmpCluster.push_back(i);  // This observation "i" is shared
                    // between both clusters
                }
            }
            m_lastPartitionSet = partitions;
        }

        printf(
            "Partitions: %u\n",
            static_cast<unsigned>(m_lastPartitionSet.size()));
    }

    MRPT_END
}

/** Return the last odometry, as a pose increment.
 */
CPose3DQuat CRangeBearingKFSLAM::getIncrementFromOdometry() const
{
    CActionRobotMovement2D::Ptr actMov2D =
        m_action->getBestMovementEstimation();
    CActionRobotMovement3D::Ptr actMov3D =
        m_action->getActionByClass<CActionRobotMovement3D>();
    if (actMov3D && !options.force_ignore_odometry)
    {
        return CPose3DQuat(actMov3D->poseChange.mean);
    }
    else if (actMov2D && !options.force_ignore_odometry)
    {
        CPose2D estMovMean;
        actMov2D->poseChange->getMean(estMovMean);
        return CPose3DQuat(CPose3D(estMovMean));
    }
    else
    {
        return CPose3DQuat();
    }
}

/** Must return the action vector u.
 * \param out_u The action vector which will be passed to OnTransitionModel
 */
void CRangeBearingKFSLAM::OnGetAction(KFArray_ACT& u) const
{
    // Get odometry estimation:
    const CPose3DQuat theIncr = getIncrementFromOdometry();

    for (KFArray_ACT::Index i = 0;
         i < static_cast<KFArray_ACT::Index>(u.size()); i++)
        u[i] = theIncr[i];
}

/** This virtual function musts implement the prediction model of the Kalman
 * filter.
 */
void CRangeBearingKFSLAM::OnTransitionModel(
    const KFArray_ACT& u, KFArray_VEH& xv, bool& out_skipPrediction) const
{
    MRPT_START

    // Do not update the vehicle pose & its covariance until we have some
    // landmarks in the map,
    // otherwise, we are imposing a lower bound to the best uncertainty from now
    // on:
    if (size_t(m_xkk.size()) == get_vehicle_size())
    {
        out_skipPrediction = true;
    }

    // Current pose: copy xyz+quat
    CPose3DQuat robotPose = getCurrentRobotPoseMean();

    // Increment pose: copy xyz+quat (explicitly unroll the loop)
    CPose3DQuat odoIncrement(UNINITIALIZED_QUATERNION);
    odoIncrement.m_coords[0] = u[0];
    odoIncrement.m_coords[1] = u[1];
    odoIncrement.m_coords[2] = u[2];
    odoIncrement.m_quat[0] = u[3];
    odoIncrement.m_quat[1] = u[4];
    odoIncrement.m_quat[2] = u[5];
    odoIncrement.m_quat[3] = u[6];

    // Pose composition:
    robotPose += odoIncrement;

    // Output:
    for (size_t i = 0; i < xv.SizeAtCompileTime; i++) xv[i] = robotPose[i];

    MRPT_END
}

/** This virtual function musts calculate the Jacobian F of the prediction
 * model.
 */
void CRangeBearingKFSLAM::OnTransitionJacobian(KFMatrix_VxV& F) const
{
    MRPT_START

    // Current pose: copy xyz+quat
    CPose3DQuat robotPose = getCurrentRobotPoseMean();

    // Odometry:
    const CPose3DQuat theIncr = getIncrementFromOdometry();

    // Compute jacobians:
    CMatrixDouble77 df_du(UNINITIALIZED_MATRIX);

    CPose3DQuatPDF::jacobiansPoseComposition(
        robotPose,  // x
        theIncr,  // u
        F,  // df_dx,
        df_du);

    MRPT_END
}

/** This virtual function musts calculate de noise matrix of the prediction
 * model.
 */
void CRangeBearingKFSLAM::OnTransitionNoise(KFMatrix_VxV& Q) const
{
    MRPT_START

    // The uncertainty of the 2D odometry, projected from the current position:
    CActionRobotMovement2D::Ptr act2D = m_action->getBestMovementEstimation();
    CActionRobotMovement3D::Ptr act3D =
        m_action->getActionByClass<CActionRobotMovement3D>();

    if (act3D && act2D)
        THROW_EXCEPTION("Both 2D & 3D odometry are present!?!?");

    CPose3DQuatPDFGaussian odoIncr;

    if ((!act3D && !act2D) || options.force_ignore_odometry)
    {
        // Use constant Q:
        Q.setZero();
        ASSERT_(size_t(options.stds_Q_no_odo.size()) == size_t(Q.cols()));
        for (size_t i = 0; i < get_vehicle_size(); i++)
            Q(i, i) = square(options.stds_Q_no_odo[i]);
        return;
    }
    else
    {
        if (act2D)
        {
            // 2D odometry:
            CPosePDFGaussian odoIncr2D;
            odoIncr2D.copyFrom(*act2D->poseChange);
            odoIncr = CPose3DQuatPDFGaussian(CPose3DPDFGaussian(odoIncr2D));
        }
        else
        {
            // 3D odometry:
            odoIncr = CPose3DQuatPDFGaussian(act3D->poseChange);
        }
    }

    odoIncr.cov(2, 2) += square(options.std_odo_z_additional);

    // Current pose: copy xyz+quat
    CPose3DQuat robotPose = getCurrentRobotPoseMean();

    // Transform from odometry increment to "relative to the robot":
    odoIncr.changeCoordinatesReference(robotPose);

    Q = odoIncr.cov;

    MRPT_END
}

void CRangeBearingKFSLAM::OnObservationModel(
    const std::vector<size_t>& idx_landmarks_to_predict,
    vector_KFArray_OBS& out_predictions) const
{
    MRPT_START

    // Mean of the prior of the robot pose:
    CPose3DQuat robotPose = getCurrentRobotPoseMean();

    // Get the sensor pose relative to the robot:
    CObservationBearingRange::Ptr obs =
        m_SF->getObservationByClass<CObservationBearingRange>();
    ASSERTMSG_(
        obs,
        "*ERROR*: This method requires an observation of type "
        "CObservationBearingRange");
    const CPose3DQuat sensorPoseOnRobot =
        CPose3DQuat(obs->sensorLocationOnRobot);

    /* -------------------------------------------
       Equations, obtained using matlab, of the relative 3D position of a
      landmark (xi,yi,zi), relative
          to a robot 6D pose (x0,y0,z0,y,p,r)
        Refer to technical report "6D EKF derivation...", 2008

        x0 y0 z0 y p r         % Robot's 6D pose
        x0s y0s z0s ys ps rs   % Sensor's 6D pose relative to robot
        xi yi zi % Absolute 3D landmark coordinates:

        Hx : dh_dxv   -> Jacobian of the observation model wrt the robot pose
        Hy : dh_dyi   -> Jacobian of the observation model wrt each landmark
      mean position

        Sizes:
         h:  Lx3
         Hx: 3Lx6
         Hy: 3Lx3

          L=# of landmarks in the map (ALL OF THEM)
      ------------------------------------------- */

    const size_t vehicle_size = get_vehicle_size();
    // const size_t  obs_size  = get_observation_size();
    const size_t feature_size = get_feature_size();

    const CPose3DQuat sensorPoseAbs = robotPose + sensorPoseOnRobot;

    const size_t N = idx_landmarks_to_predict.size();
    out_predictions.resize(N);
    for (size_t i = 0; i < N; i++)
    {
        const size_t row_in = feature_size * idx_landmarks_to_predict[i];

        // Landmark absolute 3D position in the map:
        const TPoint3D mapEst(
            m_xkk[vehicle_size + row_in + 0], m_xkk[vehicle_size + row_in + 1],
            m_xkk[vehicle_size + row_in + 2]);

        // Generate Range, yaw, pitch
        // ---------------------------------------------------
        sensorPoseAbs.sphericalCoordinates(
            mapEst,
            out_predictions[i][0],  // range
            out_predictions[i][1],  // yaw
            out_predictions[i][2]  // pitch
        );
    }

    MRPT_END
}

void CRangeBearingKFSLAM::OnObservationJacobians(
    size_t idx_landmark_to_predict, KFMatrix_OxV& Hx, KFMatrix_OxF& Hy) const
{
    MRPT_START

    // Mean of the prior of the robot pose:
    const CPose3DQuat robotPose = getCurrentRobotPoseMean();

    // Get the sensor pose relative to the robot:
    CObservationBearingRange::Ptr obs =
        m_SF->getObservationByClass<CObservationBearingRange>();
    ASSERTMSG_(
        obs,
        "*ERROR*: This method requires an observation of type "
        "CObservationBearingRange");
    const CPose3DQuat sensorPoseOnRobot =
        CPose3DQuat(obs->sensorLocationOnRobot);

    const size_t vehicle_size = get_vehicle_size();
    // const size_t  obs_size  = get_observation_size();
    const size_t feature_size = get_feature_size();

    // Compute the jacobians, needed below:
    // const CPose3DQuat  sensorPoseAbs= robotPose + sensorPoseOnRobot;
    CPose3DQuat sensorPoseAbs(UNINITIALIZED_QUATERNION);
    CMatrixFixed<kftype, 7, 7> H_senpose_vehpose(UNINITIALIZED_MATRIX);
    CMatrixFixed<kftype, 7, 7> H_senpose_senrelpose(
        UNINITIALIZED_MATRIX);  // Not actually used

    CPose3DQuatPDF::jacobiansPoseComposition(
        robotPose, sensorPoseOnRobot, H_senpose_vehpose, H_senpose_senrelpose,
        &sensorPoseAbs);

    const size_t row_in = feature_size * idx_landmark_to_predict;

    // Landmark absolute 3D position in the map:
    const TPoint3D mapEst(
        m_xkk[vehicle_size + row_in + 0], m_xkk[vehicle_size + row_in + 1],
        m_xkk[vehicle_size + row_in + 2]);

    // The Jacobian wrt the sensor pose must be transformed later on:
    KFMatrix_OxV Hx_sensor;
    double obsData[3];
    sensorPoseAbs.sphericalCoordinates(
        mapEst,
        obsData[0],  // range
        obsData[1],  // yaw
        obsData[2],  // pitch
        &Hy, &Hx_sensor);

    // Chain rule: Hx = d sensorpose / d vehiclepose   * Hx_sensor
    Hx = Hx_sensor * H_senpose_vehpose;

    MRPT_END
}

/** This is called between the KF prediction step and the update step, and the
 * application must return the observations and, when applicable, the data
 * association between these observations and the current map.
 *
 * \param out_z N vectors, each for one "observation" of length OBS_SIZE, N
 * being the number of "observations": how many observed landmarks for a map, or
 * just one if not applicable.
 * \param out_data_association An empty vector or, where applicable, a vector
 * where the i'th element corresponds to the position of the observation in the
 * i'th row of out_z within the system state vector (in the range
 * [0,getNumberOfLandmarksInTheMap()-1]), or -1 if it is a new map element and
 * we want to insert it at the end of this KF iteration.
 * \param in_S The full covariance matrix of the observation predictions (i.e.
 * the "innovation covariance matrix"). This is a M·O x M·O matrix with M=length
 * of "in_lm_indices_in_S".
 * \param in_lm_indices_in_S The indices of the map landmarks (range
 * [0,getNumberOfLandmarksInTheMap()-1]) that can be found in the matrix in_S.
 *
 *  This method will be called just once for each complete KF iteration.
 * \note It is assumed that the observations are independent, i.e. there are NO
 * cross-covariances between them.
 */
void CRangeBearingKFSLAM::OnGetObservationsAndDataAssociation(
    vector_KFArray_OBS& Z, std::vector<int>& data_association,
    const vector_KFArray_OBS& all_predictions, const KFMatrix& S,
    const std::vector<size_t>& lm_indices_in_S, const KFMatrix_OxO& R)
{
    MRPT_START

    // static const size_t vehicle_size = get_vehicle_size();
    constexpr size_t obs_size = get_observation_size();

    // Z: Observations
    CObservationBearingRange::TMeasurementList::const_iterator itObs;

    CObservationBearingRange::Ptr obs =
        m_SF->getObservationByClass<CObservationBearingRange>();
    ASSERTMSG_(
        obs,
        "*ERROR*: This method requires an observation of type "
        "CObservationBearingRange");

    const size_t N = obs->sensedData.size();
    Z.resize(N);

    size_t row;
    for (row = 0, itObs = obs->sensedData.begin();
         itObs != obs->sensedData.end(); ++itObs, ++row)
    {
        // Fill one row in Z:
        Z[row][0] = itObs->range;
        Z[row][1] = itObs->yaw;
        Z[row][2] = itObs->pitch;
    }

    // Data association:
    // ---------------------
    data_association.assign(N, -1);  // Initially, all new landmarks

    // For each observed LM:
    std::vector<size_t> obs_idxs_needing_data_assoc;
    obs_idxs_needing_data_assoc.reserve(N);

    {
        std::vector<int>::iterator itDA;
        for (row = 0, itObs = obs->sensedData.begin(),
            itDA = data_association.begin();
             itObs != obs->sensedData.end(); ++itObs, ++itDA, ++row)
        {
            // Fill data asociation: Using IDs!
            if (itObs->landmarkID < 0)
                obs_idxs_needing_data_assoc.push_back(row);
            else
            {
                mrpt::containers::bimap<
                    CLandmark::TLandmarkID, unsigned int>::iterator itID;
                if ((itID = m_IDs.find_key(itObs->landmarkID)) != m_IDs.end())
                    *itDA = itID->second;  // This row in Z corresponds to the
                // i'th map element in the state
                // vector:
            }
        }
    }

    // ---- Perform data association ----
    //  Only for observation indices in "obs_idxs_needing_data_assoc"
    if (obs_idxs_needing_data_assoc.empty())
    {
        // We don't need to do DA:
        m_last_data_association = TDataAssocInfo();

        // Save them for the case the external user wants to access them:
        for (size_t idxObs = 0; idxObs < data_association.size(); idxObs++)
        {
            int da = data_association[idxObs];
            if (da >= 0)
                m_last_data_association.results.associations[idxObs] =
                    data_association[idxObs];
        }
    }
    else
    {
        // Build a Z matrix with the observations that need dat.assoc:
        const size_t nObsDA = obs_idxs_needing_data_assoc.size();

        CMatrixDynamic<kftype> Z_obs_means(nObsDA, obs_size);
        for (size_t i = 0; i < nObsDA; i++)
        {
            const size_t idx = obs_idxs_needing_data_assoc[i];
            for (unsigned k = 0; k < obs_size; k++)
                Z_obs_means(i, k) = Z[idx][k];
        }

        // Vehicle uncertainty
        // KFMatrix_VxV Pxx = m_pkk.extractMatrix<7, 7>(0, 0);

        // Build predictions:
        // ---------------------------
        const size_t nPredictions = lm_indices_in_S.size();
        m_last_data_association.clear();

        // S is the covariance of the predictions:
        m_last_data_association.Y_pred_covs = S;

        // The means:
        m_last_data_association.Y_pred_means.setSize(nPredictions, obs_size);
        for (size_t q = 0; q < nPredictions; q++)
        {
            const size_t i = lm_indices_in_S[q];
            for (size_t w = 0; w < obs_size; w++)
                m_last_data_association.Y_pred_means(q, w) =
                    all_predictions[i][w];
            m_last_data_association.predictions_IDs.push_back(
                i);  // for the conversion of indices...
        }

        // Do Dat. Assoc :
        // ---------------------------
        if (nPredictions)
        {
            // CMatrixDouble Z_obs_cov = CMatrixDouble(R);

            mrpt::slam::data_association_full_covariance(
                Z_obs_means, m_last_data_association.Y_pred_means,
                m_last_data_association.Y_pred_covs,
                m_last_data_association.results, options.data_assoc_method,
                options.data_assoc_metric, options.data_assoc_IC_chi2_thres,
                true,  // Use KD-tree
                m_last_data_association.predictions_IDs,
                options.data_assoc_IC_metric,
                options.data_assoc_IC_ml_threshold);

            // Return pairings to the main KF algorithm:
            for (auto it = m_last_data_association.results.associations.begin();
                 it != m_last_data_association.results.associations.end(); ++it)
                data_association[it->first] = it->second;
        }
    }
    // ---- End of data association ----

    MRPT_END
}

/** This virtual function musts normalize the state vector and covariance matrix
 * (only if its necessary).
 */
void CRangeBearingKFSLAM::OnNormalizeStateVector()
{
    MRPT_START

    // m_xkk[3:6] must be a normalized quaternion:
    const double T = std::sqrt(
        square(m_xkk[3]) + square(m_xkk[4]) + square(m_xkk[5]) +
        square(m_xkk[6]));
    ASSERTMSG_(T > 0, "Vehicle pose quaternion norm is not >0!!");

    const double T_ = (m_xkk[3] < 0 ? -1.0 : 1.0) / T;  // qr>=0
    m_xkk[3] *= T_;
    m_xkk[4] *= T_;
    m_xkk[5] *= T_;
    m_xkk[6] *= T_;

    MRPT_END
}

/*---------------------------------------------------------------
                        loadOptions
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::loadOptions(const mrpt::config::CConfigFileBase& ini)
{
    // Main
    options.loadFromConfigFile(ini, "RangeBearingKFSLAM");
    KF_options.loadFromConfigFile(ini, "RangeBearingKFSLAM_KalmanFilter");
    // partition algorithm:
    mapPartitioner.options.loadFromConfigFile(ini, "RangeBearingKFSLAM");
}

/*---------------------------------------------------------------
                        loadOptions
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::TOptions::loadFromConfigFile(
    const mrpt::config::CConfigFileBase& source, const std::string& section)
{
    source.read_vector(section, "stds_Q_no_odo", stds_Q_no_odo, stds_Q_no_odo);
    ASSERT_(stds_Q_no_odo.size() == 7);
    std_sensor_range =
        source.read_float(section, "std_sensor_range", std_sensor_range);
    std_sensor_yaw = DEG2RAD(source.read_float(
        section, "std_sensor_yaw_deg", RAD2DEG(std_sensor_yaw)));
    std_sensor_pitch = DEG2RAD(source.read_float(
        section, "std_sensor_pitch_deg", RAD2DEG(std_sensor_pitch)));

    std_odo_z_additional = source.read_float(
        section, "std_odo_z_additional", std_odo_z_additional);

    MRPT_LOAD_CONFIG_VAR(doPartitioningExperiment, bool, source, section);
    MRPT_LOAD_CONFIG_VAR(partitioningMethod, int, source, section);

    MRPT_LOAD_CONFIG_VAR(create_simplemap, bool, source, section);

    MRPT_LOAD_CONFIG_VAR(force_ignore_odometry, bool, source, section);

    data_assoc_method = source.read_enum<TDataAssociationMethod>(
        section, "data_assoc_method", data_assoc_method);
    data_assoc_metric = source.read_enum<TDataAssociationMetric>(
        section, "data_assoc_metric", data_assoc_metric);
    data_assoc_IC_metric = source.read_enum<TDataAssociationMetric>(
        section, "data_assoc_IC_metric", data_assoc_IC_metric);

    MRPT_LOAD_CONFIG_VAR(data_assoc_IC_chi2_thres, double, source, section);
    MRPT_LOAD_CONFIG_VAR(data_assoc_IC_ml_threshold, double, source, section);

    MRPT_LOAD_CONFIG_VAR(quantiles_3D_representation, float, source, section);
}

/*---------------------------------------------------------------
                        Constructor
  ---------------------------------------------------------------*/
CRangeBearingKFSLAM::TOptions::TOptions()
    : stds_Q_no_odo(get_vehicle_size(), 0),

      std_sensor_yaw(DEG2RAD(0.2f)),
      std_sensor_pitch(DEG2RAD(0.2f))

{
    stds_Q_no_odo[0] = stds_Q_no_odo[1] = stds_Q_no_odo[2] = 0.10f;
    stds_Q_no_odo[3] = stds_Q_no_odo[4] = stds_Q_no_odo[5] = stds_Q_no_odo[6] =
        0.05f;
}

/*---------------------------------------------------------------
                        dumpToTextStream
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::TOptions::dumpToTextStream(std::ostream& out) const
{
    using namespace mrpt::typemeta;

    out << "\n----------- [CRangeBearingKFSLAM::TOptions] ------------ \n\n";

    out << mrpt::format(
        "doPartitioningExperiment                = %c\n",
        doPartitioningExperiment ? 'Y' : 'N');
    out << mrpt::format(
        "partitioningMethod                      = %i\n", partitioningMethod);
    out << mrpt::format(
        "data_assoc_method                       = %s\n",
        TEnumType<TDataAssociationMethod>::value2name(data_assoc_method)
            .c_str());
    out << mrpt::format(
        "data_assoc_metric                       = %s\n",
        TEnumType<TDataAssociationMetric>::value2name(data_assoc_metric)
            .c_str());
    out << mrpt::format(
        "data_assoc_IC_chi2_thres                = %.06f\n",
        data_assoc_IC_chi2_thres);
    out << mrpt::format(
        "data_assoc_IC_metric                    = %s\n",
        TEnumType<TDataAssociationMetric>::value2name(data_assoc_IC_metric)
            .c_str());
    out << mrpt::format(
        "data_assoc_IC_ml_threshold              = %.06f\n",
        data_assoc_IC_ml_threshold);

    out << "\n";
}

void CRangeBearingKFSLAM::OnInverseObservationModel(
    const KFArray_OBS& in_z, KFArray_FEAT& yn, KFMatrix_FxV& dyn_dxv,
    KFMatrix_FxO& dyn_dhn) const
{
    MRPT_START

    CObservationBearingRange::Ptr obs =
        m_SF->getObservationByClass<CObservationBearingRange>();
    ASSERTMSG_(
        obs,
        "*ERROR*: This method requires an observation of type "
        "CObservationBearingRange");
    const CPose3DQuat sensorPoseOnRobot =
        CPose3DQuat(obs->sensorLocationOnRobot);

    // Mean of the prior of the robot pose:
    const CPose3DQuat robotPose = getCurrentRobotPoseMean();

    // const CPose3DQuat  sensorPoseAbs= robotPose + sensorPoseOnRobot;
    CPose3DQuat sensorPoseAbs(UNINITIALIZED_QUATERNION);
    CMatrixFixed<kftype, 7, 7> dsensorabs_dvehpose(UNINITIALIZED_MATRIX);
    CMatrixFixed<kftype, 7, 7> dsensorabs_dsenrelpose(
        UNINITIALIZED_MATRIX);  // Not actually used

    CPose3DQuatPDF::jacobiansPoseComposition(
        robotPose, sensorPoseOnRobot, dsensorabs_dvehpose,
        dsensorabs_dsenrelpose, &sensorPoseAbs);

    kftype hn_r = in_z[0];
    kftype hn_y = in_z[1];
    kftype hn_p = in_z[2];

    kftype chn_y = cos(hn_y);
    kftype shn_y = sin(hn_y);
    kftype chn_p = cos(hn_p);
    kftype shn_p = sin(hn_p);

    /* -------------------------------------------
        syms H h_range h_yaw h_pitch real;
        H=[ h_range ; h_yaw ; h_pitch ];

        syms X xi_ yi_ zi_ real;
        xi_ = h_range * cos(h_yaw) * cos(h_pitch);
        yi_ = h_range * sin(h_yaw) * cos(h_pitch);
        zi_ = -h_range * sin(h_pitch);

        X=[xi_ yi_ zi_];
        jacob_inv_mod=jacobian(X,H)
      ------------------------------------------- */

    // The new point, relative to the sensor:
    const TPoint3D yn_rel_sensor(
        hn_r * chn_y * chn_p, hn_r * shn_y * chn_p, -hn_r * shn_p);

    // The Jacobian of the 3D point in the coordinate system of the sensor:
    /*
        [ cos(h_pitch)*cos(h_yaw), -h_range*cos(h_pitch)*sin(h_yaw),
       -h_range*cos(h_yaw)*sin(h_pitch)]
        [ cos(h_pitch)*sin(h_yaw),  h_range*cos(h_pitch)*cos(h_yaw),
       -h_range*sin(h_pitch)*sin(h_yaw)]
        [           -sin(h_pitch),                                0,
       -h_range*cos(h_pitch)]
    */
    const kftype values_dynlocal_dhn[] = {chn_p * chn_y,
                                          -hn_r * chn_p * shn_y,
                                          -hn_r * chn_y * shn_p,
                                          chn_p * shn_y,
                                          hn_r * chn_p * chn_y,
                                          -hn_r * shn_p * shn_y,
                                          -shn_p,
                                          0,
                                          -hn_r * chn_p};
    const KFMatrix_FxO dynlocal_dhn(values_dynlocal_dhn);

    KFMatrix_FxF jacob_dyn_dynrelsensor(UNINITIALIZED_MATRIX);
    KFMatrix_FxV jacob_dyn_dsensorabs(UNINITIALIZED_MATRIX);

    sensorPoseAbs.composePoint(
        yn_rel_sensor.x, yn_rel_sensor.y,
        yn_rel_sensor.z,  // yn rel. to the sensor
        yn[0], yn[1], yn[2],  // yn in global coords
        &jacob_dyn_dynrelsensor, &jacob_dyn_dsensorabs);

    dyn_dhn = jacob_dyn_dynrelsensor * dynlocal_dhn;

    // dyn_dxv =
    dyn_dxv = jacob_dyn_dsensorabs * dsensorabs_dvehpose;

    MRPT_END
}

/** If applicable to the given problem, do here any special handling of adding a
 * new landmark to the map.
 * \param in_obsIndex The index of the observation whose inverse sensor is to
 * be computed. It corresponds to the row in in_z where the observation can be
 * found.
 * \param in_idxNewFeat The index that this new feature will have in the state
 * vector (0:just after the vehicle state, 1: after that,...). Save this number
 * so data association can be done according to these indices.
 * \sa OnInverseObservationModel
 */
void CRangeBearingKFSLAM::OnNewLandmarkAddedToMap(
    const size_t in_obsIdx, const size_t in_idxNewFeat)
{
    MRPT_START

    CObservationBearingRange::Ptr obs =
        m_SF->getObservationByClass<CObservationBearingRange>();
    ASSERTMSG_(
        obs,
        "*ERROR*: This method requires an observation of type "
        "CObservationBearingRange");

    // ----------------------------------------------
    // introduce in the lists of ID<->index in map:
    // ----------------------------------------------
    ASSERT_(in_obsIdx < obs->sensedData.size());
    if (obs->sensedData[in_obsIdx].landmarkID >= 0)
    {
        // The sensor provides us a LM ID... use it:
        m_IDs.insert(obs->sensedData[in_obsIdx].landmarkID, in_idxNewFeat);
    }
    else
    {
        // Features do not have IDs... use indices:
        m_IDs.insert(in_idxNewFeat, in_idxNewFeat);
    }

    MRPT_END
}

/*---------------------------------------------------------------
                        getAs3DObject
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::getAs3DObject(
    mrpt::opengl::CSetOfObjects::Ptr& outObj) const
{
    outObj->clear();

    // ------------------------------------------------
    //  Add the XYZ corner for the current area:
    // ------------------------------------------------
    outObj->insert(opengl::stock_objects::CornerXYZ());

    // 3D ellipsoid for robot pose:
    CPointPDFGaussian pointGauss;
    pointGauss.mean.x(m_xkk[0]);
    pointGauss.mean.y(m_xkk[1]);
    pointGauss.mean.z(m_xkk[2]);
    pointGauss.cov = m_pkk.blockCopy<3, 3>(0, 0);

    {
        auto ellip = opengl::CEllipsoid3D::Create();

        ellip->setPose(pointGauss.mean);
        ellip->setCovMatrix(pointGauss.cov);
        ellip->enableDrawSolid3D(false);
        ellip->setQuantiles(options.quantiles_3D_representation);
        ellip->set3DsegmentsCount(10);
        ellip->setColor(1, 0, 0);

        outObj->insert(ellip);
    }

    // 3D ellipsoids for landmarks:
    const size_t nLMs = this->getNumberOfLandmarksInTheMap();
    for (size_t i = 0; i < nLMs; i++)
    {
        pointGauss.mean.x(
            m_xkk[get_vehicle_size() + get_feature_size() * i + 0]);
        pointGauss.mean.y(
            m_xkk[get_vehicle_size() + get_feature_size() * i + 1]);
        pointGauss.mean.z(
            m_xkk[get_vehicle_size() + get_feature_size() * i + 2]);

        pointGauss.cov =
            m_pkk.blockCopy<get_feature_size(), get_feature_size()>(
                get_vehicle_size() + get_feature_size() * i,
                get_vehicle_size() + get_feature_size() * i);

        auto ellip = opengl::CEllipsoid3D::Create();

        ellip->setName(format("%u", static_cast<unsigned int>(i)));
        ellip->enableShowName(true);
        ellip->setPose(pointGauss.mean);
        ellip->setCovMatrix(pointGauss.cov);
        ellip->enableDrawSolid3D(false);
        ellip->setQuantiles(options.quantiles_3D_representation);
        ellip->set3DsegmentsCount(10);

        ellip->setColor(0, 0, 1);

        // Set color depending on partitions?
        if (options.doPartitioningExperiment)
        {
            // This is done for each landmark:
            map<int, bool> belongToPartition;
            const CSimpleMap* SFs =
                &m_SFs;  // mapPartitioner.getSequenceOfFrames();

            for (size_t p = 0; p < m_lastPartitionSet.size(); p++)
            {
                for (size_t w = 0; w < m_lastPartitionSet[p].size(); w++)
                {
                    // Check if landmark #i is in the SF of
                    // m_lastPartitionSet[p][w]:
                    CLandmark::TLandmarkID i_th_ID = m_IDs.inverse(i);

                    // Look for the lm_ID in the SF:
                    CPose3DPDF::Ptr pdf;
                    CSensoryFrame::Ptr SF_i;
                    SFs->get(m_lastPartitionSet[p][w], pdf, SF_i);

                    CObservationBearingRange::Ptr obs =
                        SF_i->getObservationByClass<CObservationBearingRange>();

                    for (auto& o : obs->sensedData)
                    {
                        if (o.landmarkID == i_th_ID)
                        {
                            belongToPartition[p] = true;
                            break;
                        }
                    }  // end for o
                }  // end for w
            }  // end for p

            // Build label:
            string strParts("[");

            for (auto it = belongToPartition.begin();
                 it != belongToPartition.end(); ++it)
            {
                if (it != belongToPartition.begin()) strParts += string(",");
                strParts += format("%i", it->first);
            }

            ellip->setName(ellip->getName() + strParts + string("]"));
        }

        outObj->insert(ellip);
    }
}

/*---------------------------------------------------------------
                        getLastPartitionLandmarks
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::getLastPartitionLandmarksAsIfFixedSubmaps(
    size_t K, std::vector<std::vector<uint32_t>>& landmarksMembership)
{
    // temporary copy:
    std::vector<std::vector<uint32_t>> tmpParts = m_lastPartitionSet;

    // Fake "m_lastPartitionSet":

    // Fixed partitions every K observations:
    vector<std::vector<uint32_t>> partitions;
    std::vector<uint32_t> tmpCluster;

    for (size_t i = 0; i < m_SFs.size(); i++)
    {
        tmpCluster.push_back(i);
        if ((i % K) == 0)
        {
            partitions.push_back(tmpCluster);
            tmpCluster.clear();
            tmpCluster.push_back(
                i);  // This observation "i" is shared between both clusters
        }
    }
    m_lastPartitionSet = partitions;

    // Call the actual method:
    getLastPartitionLandmarks(landmarksMembership);

    // Replace copy:
    m_lastPartitionSet = tmpParts;  //-V519
}

/*---------------------------------------------------------------
                        getLastPartitionLandmarks
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::getLastPartitionLandmarks(
    std::vector<std::vector<uint32_t>>& landmarksMembership) const
{
    landmarksMembership.clear();

    // All the computation is made based on "m_lastPartitionSet"

    if (!options.doPartitioningExperiment) return;

    const size_t nLMs = this->getNumberOfLandmarksInTheMap();
    for (size_t i = 0; i < nLMs; i++)
    {
        map<int, bool> belongToPartition;
        const CSimpleMap* SFs =
            &m_SFs;  // mapPartitioner.getSequenceOfFrames();

        for (size_t p = 0; p < m_lastPartitionSet.size(); p++)
        {
            for (size_t w = 0; w < m_lastPartitionSet[p].size(); w++)
            {
                // Check if landmark #i is in the SF of
                // m_lastPartitionSet[p][w]:
                CLandmark::TLandmarkID i_th_ID = m_IDs.inverse(i);

                // Look for the lm_ID in the SF:
                CPose3DPDF::Ptr pdf;
                CSensoryFrame::Ptr SF_i;
                SFs->get(m_lastPartitionSet[p][w], pdf, SF_i);

                CObservationBearingRange::Ptr obs =
                    SF_i->getObservationByClass<CObservationBearingRange>();

                for (auto& o : obs->sensedData)
                {
                    if (o.landmarkID == i_th_ID)
                    {
                        belongToPartition[p] = true;
                        break;
                    }
                }  // end for o
            }  // end for w
        }  // end for p

        // Build membership list:
        std::vector<uint32_t> membershipOfThisLM;

        for (auto& it : belongToPartition)
            membershipOfThisLM.push_back(it.first);

        landmarksMembership.push_back(membershipOfThisLM);
    }  // end for i
}

/*---------------------------------------------------------------
            computeOffDiagonalBlocksApproximationError
  ---------------------------------------------------------------*/
double CRangeBearingKFSLAM::computeOffDiagonalBlocksApproximationError(
    const std::vector<std::vector<uint32_t>>& landmarksMembership) const
{
    MRPT_START

    // Compute the information matrix:
    CMatrixDynamic<kftype> fullCov(m_pkk);
    size_t i;
    for (i = 0; i < get_vehicle_size(); i++)
        fullCov(i, i) = max(fullCov(i, i), 1e-6);

    CMatrixDynamic<kftype> H(fullCov.inverse_LLt());
    H.array().abs();  // Replace by absolute values:

    double sumOffBlocks = 0;
    unsigned int nLMs = landmarksMembership.size();

    ASSERT_(
        int(get_vehicle_size() + nLMs * get_feature_size()) == fullCov.cols());

    for (i = 0; i < nLMs; i++)
    {
        for (size_t j = i + 1; j < nLMs; j++)
        {
            // Sum the cross cov. between LM(i) and LM(j)??
            // --> Only if there is no common cluster:
            if (0 == math::countCommonElements(
                         landmarksMembership[i], landmarksMembership[j]))
            {
                size_t col = get_vehicle_size() + i * get_feature_size();
                size_t row = get_vehicle_size() + j * get_feature_size();
                sumOffBlocks += 2 * H.block<2, 2>(row, col).sum();
            }
        }
    }

    return sumOffBlocks / H.asEigen()
                              .block(
                                  get_vehicle_size(), get_vehicle_size(),
                                  H.rows() - get_vehicle_size(),
                                  H.cols() - get_vehicle_size())
                              .sum();  // Starting (7,7)-end
    MRPT_END
}

/*---------------------------------------------------------------
              reconsiderPartitionsNow
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::reconsiderPartitionsNow()
{
    // A different buffer for making this
    // thread-safe some day...
    vector<std::vector<uint32_t>> partitions;
    mapPartitioner.updatePartitions(partitions);
    m_lastPartitionSet = partitions;
}

/*---------------------------------------------------------------
              saveMapAndPathRepresentationAsMATLABFile
  ---------------------------------------------------------------*/
void CRangeBearingKFSLAM::saveMapAndPath2DRepresentationAsMATLABFile(
    const string& fil, float stdCount, const string& styleLandmarks,
    const string& stylePath, const string& styleRobot) const
{
    FILE* f = os::fopen(fil.c_str(), "wt");
    if (!f) return;

    CMatrixDouble cov(2, 2);
    CVectorDouble mean(2);

    // Header:
    os::fprintf(
        f,
        "%%--------------------------------------------------------------------"
        "\n");
    os::fprintf(f, "%% File automatically generated using the MRPT method:\n");
    os::fprintf(
        f,
        "%% "
        "'CRangeBearingKFSLAM::saveMapAndPath2DRepresentationAsMATLABFile'\n");
    os::fprintf(f, "%%\n");
    os::fprintf(f, "%%                        ~ MRPT ~\n");
    os::fprintf(
        f, "%%  Jose Luis Blanco Claraco, University of Malaga @ 2008\n");
    os::fprintf(f, "%%      https://www.mrpt.org/     \n");
    os::fprintf(
        f,
        "%%--------------------------------------------------------------------"
        "\n");

    // Main code:
    os::fprintf(f, "hold on;\n\n");

    const size_t nLMs = this->getNumberOfLandmarksInTheMap();

    for (size_t i = 0; i < nLMs; i++)
    {
        size_t idx = get_vehicle_size() + i * get_feature_size();

        cov(0, 0) = m_pkk(idx + 0, idx + 0);
        cov(1, 1) = m_pkk(idx + 1, idx + 1);
        cov(0, 1) = cov(1, 0) = m_pkk(idx + 0, idx + 1);

        mean[0] = m_xkk[idx + 0];
        mean[1] = m_xkk[idx + 1];

        // Command to draw the 2D ellipse:
        os::fprintf(
            f, "%s",
            math::MATLAB_plotCovariance2D(cov, mean, stdCount, styleLandmarks)
                .c_str());
    }

    // Now: the robot path:
    // ------------------------------
    if (m_SFs.size())
    {
        os::fprintf(f, "\nROB_PATH=[");
        for (size_t i = 0; i < m_SFs.size(); i++)
        {
            CSensoryFrame::Ptr dummySF;
            CPose3DPDF::Ptr pdf3D;
            m_SFs.get(i, pdf3D, dummySF);

            CPose3D p;
            pdf3D->getMean(p);

            os::fprintf(f, "%.04f %.04f", p.x(), p.y());
            if (i < (m_SFs.size() - 1)) os::fprintf(f, ";");
        }
        os::fprintf(f, "];\n");

        os::fprintf(
            f, "plot(ROB_PATH(:,1),ROB_PATH(:,2),'%s');\n", stylePath.c_str());
    }

    // The robot pose:
    cov(0, 0) = m_pkk(0, 0);
    cov(1, 1) = m_pkk(1, 1);
    cov(0, 1) = cov(1, 0) = m_pkk(0, 1);

    mean[0] = m_xkk[0];
    mean[1] = m_xkk[1];

    os::fprintf(
        f, "%s",
        math::MATLAB_plotCovariance2D(cov, mean, stdCount, styleRobot).c_str());

    os::fprintf(f, "\naxis equal;\n");
    os::fclose(f);
}

/** Computes A=A-B, which may need to be re-implemented depending on the
 * topology of the individual scalar components (eg, angles).
 */
void CRangeBearingKFSLAM::OnSubstractObservationVectors(
    KFArray_OBS& A, const KFArray_OBS& B) const
{
    A -= B;
    mrpt::math::wrapToPiInPlace(A[1]);
    mrpt::math::wrapToPiInPlace(A[2]);
}

/** Return the observation NOISE covariance matrix, that is, the model of the
 * Gaussian additive noise of the sensor.
 * \param out_R The noise covariance matrix. It might be non diagonal, but
 * it'll usually be.
 */
void CRangeBearingKFSLAM::OnGetObservationNoise(KFMatrix_OxO& out_R) const
{
    out_R(0, 0) = square(options.std_sensor_range);
    out_R(1, 1) = square(options.std_sensor_yaw);
    out_R(2, 2) = square(options.std_sensor_pitch);
}

/** This will be called before OnGetObservationsAndDataAssociation to allow the
 * application to reduce the number of covariance landmark predictions to be
 * made.
 *  For example, features which are known to be "out of sight" shouldn't be
 * added to the output list to speed up the calculations.
 * \param in_all_prediction_means The mean of each landmark predictions; the
 * computation or not of the corresponding covariances is what we're trying to
 * determined with this method.
 * \param out_LM_indices_to_predict The list of landmark indices in the map
 * [0,getNumberOfLandmarksInTheMap()-1] that should be predicted.
 * \note This is not a pure virtual method, so it should be implemented only if
 * desired. The default implementation returns a vector with all the landmarks
 * in the map.
 * \sa OnGetObservations, OnDataAssociation
 */
void CRangeBearingKFSLAM::OnPreComputingPredictions(
    const vector_KFArray_OBS& prediction_means,
    std::vector<size_t>& out_LM_indices_to_predict) const
{
    CObservationBearingRange::Ptr obs =
        m_SF->getObservationByClass<CObservationBearingRange>();
    ASSERTMSG_(
        obs,
        "*ERROR*: This method requires an observation of type "
        "CObservationBearingRange");

#define USE_HEURISTIC_PREDICTION

#ifdef USE_HEURISTIC_PREDICTION
    const double sensor_max_range = obs->maxSensorDistance;
    const double fov_yaw = obs->fieldOfView_yaw;
    const double fov_pitch = obs->fieldOfView_pitch;

    const double max_vehicle_loc_uncertainty =
        4 * std::sqrt(m_pkk(0, 0) + m_pkk(1, 1) + m_pkk(2, 2));
#endif

    out_LM_indices_to_predict.clear();
    for (size_t i = 0; i < prediction_means.size(); i++)
    {
#ifndef USE_HEURISTIC_PREDICTION
        out_LM_indices_to_predict.push_back(i);
#else
        // Heuristic: faster but doesn't work always!
        if (prediction_means[i][0] <
                (15 + sensor_max_range + max_vehicle_loc_uncertainty +
                 4 * options.std_sensor_range) &&
            fabs(prediction_means[i][1]) <
                (30.0_deg + 0.5 * fov_yaw + 4 * options.std_sensor_yaw) &&
            fabs(prediction_means[i][2]) <
                (30.0_deg + 0.5 * fov_pitch + 4 * options.std_sensor_pitch))
        {
            out_LM_indices_to_predict.push_back(i);
        }
#endif
    }
}