CICP.h

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          |
   +------------------------------------------------------------------------+ */
#pragma once

#include <mrpt/config/CLoadableOptions.h>
#include <mrpt/slam/CMetricMapsAlignmentAlgorithm.h>
#include <mrpt/typemeta/TEnumType.h>

namespace mrpt::slam
{
/** The ICP algorithm selection, used in mrpt::slam::CICP::options  \ingroup
 * mrpt_slam_grp  */
enum TICPAlgorithm
{
    icpClassic = 0,
    icpLevenbergMarquardt
};

/** ICP covariance estimation methods, used in mrpt::slam::CICP::options
 * \ingroup mrpt_slam_grp  */
enum TICPCovarianceMethod
{
    /** Use the covariance of the optimal registration, disregarding uncertainty
       in data association */
    icpCovLinealMSE = 0,
    /** Covariance of a least-squares optimizer (includes data association
       uncertainty) */
    icpCovFiniteDifferences
};

/** Several implementations of ICP (Iterative closest point) algorithms for
 * aligning two point maps or a point map wrt a grid map.
 *
 *  CICP::AlignPDF() or CICP::Align() are the two main entry points of the
 * algorithm.
 *
 *  To choose among existing ICP algorithms or customizing their parameters, see
 * CICP::TConfigParams and the member \a options.
 *
 *  There exists an extension of the original ICP algorithm that provides
 * multihypotheses-support for the correspondences, and which generates a
 * Sum-of-Gaussians (SOG)
 *    PDF as output. See mrpt::tfest::se2_l2_robust()
 *
 * For further details on the implemented methods, check the web:
 *   https://www.mrpt.org/Iterative_Closest_Point_(ICP)_and_other_matching_algorithms
 *
 *  For a paper explaining some of the basic equations, see for example:
 *   J. Martinez, J. Gonzalez, J. Morales, A. Mandow, A. Garcia-Cerezo,
 *   "Mobile robot motion estimation by 2D scan matching with genetic and
 * iterative closest point algorithms",
 *    Journal of Field Robotics, vol. 23, no. 1, pp. 21-34, 2006. (
 * http://babel.isa.uma.es/~jlblanco/papers/martinez2006gil.pdf )
 *
 * \sa CMetricMapsAlignmentAlgorithm
 * \ingroup mrpt_slam_grp
 */
class CICP : public mrpt::slam::CMetricMapsAlignmentAlgorithm
{
   public:
    /** The ICP algorithm configuration data
     */
    class TConfigParams : public mrpt::config::CLoadableOptions
    {
       public:
        void loadFromConfigFile(
            const mrpt::config::CConfigFileBase& source,
            const std::string& section) override;  // See base docs
        void saveToConfigFile(
            mrpt::config::CConfigFileBase& c,
            const std::string& s) const override;

        /** @name Algorithms selection
            @{ */
        /** The algorithm to use (default: icpClassic). See
         * https://www.mrpt.org/tutorials/programming/scan-matching-and-icp/ for
         * details */
        TICPAlgorithm ICP_algorithm{icpClassic};
        /** The method to use for covariance estimation (Default:
         * icpCovFiniteDifferences) */
        TICPCovarianceMethod ICP_covariance_method{icpCovFiniteDifferences};
        /** @} */

        /** @name Correspondence-finding criteria
            @{ */
        bool onlyUniqueRobust{false};
        //! if this option is enabled only the closest
        //! correspondence for each reference point will
        //! be kept (default=false).
        /** @} */

        /** @name Termination criteria
            @{ */
        /** Maximum number of iterations to run. */
        unsigned int maxIterations{40};
        /** If the correction in all translation coordinates (X,Y,Z) is below
         * this threshold (in meters), iterations are terminated (Default:1e-6)
         */
        double minAbsStep_trans{1e-6};
        /** If the correction in all rotation coordinates (yaw,pitch,roll) is
         * below this threshold (in radians), iterations are terminated
         * (Default:1e-6) */
        double minAbsStep_rot{1e-6};
        /** @} */

        /** Initial threshold distance for two points to become a
         * correspondence. */
        double thresholdDist{0.75}, thresholdAng{0.15 * M_PI / 180.0};
        /** The scale factor for thresholds everytime convergence is achieved.
         */
        double ALFA{0.5};
        /** The size for threshold such that iterations will stop, since it is
         * considered precise enough. */
        double smallestThresholdDist{0.1};

        /** This is the normalization constant \f$ \sigma^2_p \f$ that is used
         * to scale the whole 3x3 covariance.
         *  This has a default value of \f$ (0.02)^2 \f$, that is, a 2cm sigma.
         *  See paper: J.L. Blanco, J. Gonzalez-Jimenez, J.A.
         * Fernandez-Madrigal, "A Robust, Multi-Hypothesis Approach to Matching
         * Occupancy Grid Maps", Robotica, vol. 31, no. 5, pp. 687-701, 2013.
         */
        double covariance_varPoints{0.02 * 0.02};

        /** Perform a RANSAC step, mrpt::tfest::se2_l2_robust(), after the ICP
         * convergence, to obtain a better estimation of the pose PDF. */
        bool doRANSAC{false};

        /** @name RANSAC-step options for mrpt::tfest::se2_l2_robust() if \a
         * doRANSAC=true
         * @{ */
        unsigned int ransac_minSetSize{3}, ransac_maxSetSize{20},
            ransac_nSimulations{100};
        double ransac_mahalanobisDistanceThreshold{3.0};
        /** RANSAC-step option: The standard deviation in X,Y of
         * landmarks/points which are being matched (used to compute covariances
         * in the SoG) */
        double normalizationStd{0.02};
        bool ransac_fuseByCorrsMatch{true};
        double ransac_fuseMaxDiffXY{0.01},
            ransac_fuseMaxDiffPhi{0.1 * M_PI / 180.0};
        /** @} */

        /** Cauchy kernel rho, for estimating the optimal transformation
         * covariance, in meters (default = 0.07m) */
        double kernel_rho{0.07};
        /** Whether to use kernel_rho to smooth distances, or use distances
         * directly (default=true) */
        bool use_kernel{true};
        /** [LM method only] The size of the perturbance in x & y used to
         * estimate the Jacobians of the square error (default=0.05) */
        double Axy_aprox_derivatives{0.05};
        /** [LM method only] The initial value of the lambda parameter in the LM
         * method (default=1e-4) */
        double LM_initial_lambda{1e-4};

        /** Skip the computation of the covariance (saves some time)
         * (default=false) */
        bool skip_cov_calculation{false};
        /** Skip the (sometimes) expensive evaluation of the term 'quality' at
         * ICP output (Default=true) */
        bool skip_quality_calculation{true};

        /** Decimation of the point cloud being registered against the reference
         * one (default=5) - set to 1 to have the older (MRPT <0.9.5) behavior
         *  of not approximating ICP by ignoring the correspondence of some
         * points. The speed-up comes from a decimation of the number of KD-tree
         * queries,
         *  the most expensive step in ICP */
        uint32_t corresponding_points_decimation{5};
    };

    /** The options employed by the ICP align. */
    TConfigParams options;

    /** Constructor with the default options */
    CICP() : options() {}
    /** Constructor that directly set the ICP params from a given struct \sa
     * options */
    CICP(const TConfigParams& icpParams) : options(icpParams) {}
    /** Destructor */
    ~CICP() override = default;
    /** The ICP algorithm return information*/
    struct TReturnInfo
    {
        TReturnInfo() = default;
        /** The number of executed iterations until convergence */
        unsigned short nIterations{0};
        /** A goodness measure for the alignment, it is a [0,1] range indicator
         * of percentage of correspondences. */
        float goodness{0};
        /** A measure of the 'quality' of the local minimum of the sqr. error
         * found by the method. Higher values are better. Low values will be
         * found in ill-conditioned situations (e.g. a corridor) */
        float quality{0};
    };

    /** An implementation of CMetricMapsAlignmentAlgorithm for the case of a
     * point maps and a occupancy grid/point map.
     *
     *  This method computes the PDF of the displacement (relative pose) between
     *   two maps: <b>the relative pose of m2 with respect to m1</b>. This pose
     *   is returned as a PDF rather than a single value.
     *
     *  \note This method can be configurated with "CICP::options"
     *  \note The output PDF is a CPosePDFGaussian if "doRANSAC=false", or a
     * CPosePDFSOG otherwise.
     *
     * \param m1            [IN] The first map (CAN BE A mrpt::poses::CPointsMap
     * derived
     * class or a mrpt::slam::COccupancyGrid2D class)
     * \param m2            [IN] The second map. (MUST BE A
     * mrpt::poses::CPointsMap
     * derived class)The pose of this map respect to m1 is to be estimated.
     * \param initialEstimationPDF  [IN] An initial gross estimation for the
     * displacement.
     * \param runningTime   [OUT] A pointer to a container for obtaining the
     * algorithm running time in seconds, or nullptr if you don't need it.
     * \param info          [OUT] A pointer to a CICP::TReturnInfo, or nullptr
     * if
     * it
     * isn't needed.
     *
     * \return A smart pointer to the output estimated pose PDF.
     *
     * \sa CMetricMapsAlignmentAlgorithm, CICP::options, CICP::TReturnInfo
     */
    mrpt::poses::CPosePDF::Ptr AlignPDF(
        const mrpt::maps::CMetricMap* m1, const mrpt::maps::CMetricMap* m2,
        const mrpt::poses::CPosePDFGaussian& initialEstimationPDF,
        float* runningTime = nullptr, void* info = nullptr) override;

    // See base class for docs
    mrpt::poses::CPose3DPDF::Ptr Align3DPDF(
        const mrpt::maps::CMetricMap* m1, const mrpt::maps::CMetricMap* m2,
        const mrpt::poses::CPose3DPDFGaussian& initialEstimationPDF,
        float* runningTime = nullptr, void* info = nullptr) override;

   protected:
    /** Computes:
     *  \f[ K(x^2) = \frac{x^2}{x^2+\rho^2}  \f]
     *  or just return the input if options.useKernel = false.
     */
    float kernel(const float& x2, const float& rho2);

    mrpt::poses::CPosePDF::Ptr ICP_Method_Classic(
        const mrpt::maps::CMetricMap* m1, const mrpt::maps::CMetricMap* m2,
        const mrpt::poses::CPosePDFGaussian& initialEstimationPDF,
        TReturnInfo& outInfo);
    mrpt::poses::CPosePDF::Ptr ICP_Method_LM(
        const mrpt::maps::CMetricMap* m1, const mrpt::maps::CMetricMap* m2,
        const mrpt::poses::CPosePDFGaussian& initialEstimationPDF,
        TReturnInfo& outInfo);
    mrpt::poses::CPose3DPDF::Ptr ICP3D_Method_Classic(
        const mrpt::maps::CMetricMap* m1, const mrpt::maps::CMetricMap* m2,
        const mrpt::poses::CPose3DPDFGaussian& initialEstimationPDF,
        TReturnInfo& outInfo);
};
}  // namespace mrpt::slam
MRPT_ENUM_TYPE_BEGIN(mrpt::slam::TICPAlgorithm)
using namespace mrpt::slam;
MRPT_FILL_ENUM(icpClassic);
MRPT_FILL_ENUM(icpLevenbergMarquardt);
MRPT_ENUM_TYPE_END()

MRPT_ENUM_TYPE_BEGIN(mrpt::slam::TICPCovarianceMethod)
using namespace mrpt::slam;
MRPT_FILL_ENUM(icpCovLinealMSE);
MRPT_FILL_ENUM(icpCovFiniteDifferences);
MRPT_ENUM_TYPE_END()