DroneKalmanFilter.h

Go to the documentation of this file.

#pragma once

#ifndef __DRONEKALMANFILTER_H
#define __DRONEKALMANFILTER_H

#include <TooN/TooN.h>
#include <deque>
#include <iostream>
#include <fstream>
#include <ardrone_autonomy/Navdata.h>
#include <geometry_msgs/Twist.h>
#include <geometry_msgs/TwistStamped.h>
#include <pthread.h>
#include "../HelperFunctions.h"
#include "uga_tum_ardrone/filter_state.h"

class EstimationNode;


class ScaleStruct
{
public:
        TooN::Vector<3> ptam;
        TooN::Vector<3> imu;
        double ptamNorm;
        double imuNorm;
        double alphaSingleEstimate;
        double pp, ii, pi;

        inline double computeEstimator(double spp, double sii, double spi, double stdDevPTAM = 0.2, double stdDevIMU = 0.1)
        {
                double sII = stdDevPTAM * stdDevPTAM * sii;
                double sPP = stdDevIMU * stdDevIMU * spp;
                double sPI = stdDevIMU * stdDevPTAM * spi;

                double tmp = (sII-sPP)*(sII-sPP) + 4*sPI*sPI;
                if(tmp <= 0) tmp = 1e-5;        // numeric issues
                return 0.5*((sII-sPP)+sqrt(tmp)) / (stdDevPTAM * stdDevPTAM * spi);

        }

        inline ScaleStruct(TooN::Vector<3> ptamDist, TooN::Vector<3> imuDist)
        {
                ptam = ptamDist;
                imu = imuDist;
                pp = ptam[0]*ptam[0] + ptam[1]*ptam[1] + ptam[2]*ptam[2];
                ii = imu[0]*imu[0] + imu[1]*imu[1] + imu[2]*imu[2];
                pi = imu[0]*ptam[0] + imu[1]*ptam[1] + imu[2]*ptam[2];

                ptamNorm = sqrt(pp);
                imuNorm = sqrt(ii);

                alphaSingleEstimate = computeEstimator(pp,ii,pi);
        }

        inline bool operator < (const ScaleStruct& comp) const
        {
                return alphaSingleEstimate < comp.alphaSingleEstimate;
        }
};


// KalmanFilter with two components (pose, speed)
class PVFilter
{
public:
        TooN::Vector<2> state;
        TooN::Matrix<2,2> var;


        // constructors
        inline PVFilter(TooN::Vector<2> state, TooN::Matrix<2,2> var)
                : state(state), var(var)
        {
        }

        inline PVFilter(double pose, double speed)
                : state(TooN::makeVector(pose,speed)), var(TooN::Zeros)
        {
        }

        inline PVFilter(double pose)
                : state(TooN::makeVector(pose,0)), var(TooN::Zeros)
        {
                var(1,1) = 1e10;
        }

        inline PVFilter()
                : state(TooN::makeVector(0,0)), var(TooN::Zeros)
        {
                var(1,1) = var(0,0) = 1e10;
        }

        // observe
        inline void observePose(double obs, double obsVar)
        {
                /* MATLAB:
                H = [1 0];
                K = (uncertainty * H') / ((H * uncertainty * H') + obsVar);
                state = state + K * (obs - H*state);
                var = (eye(2)-K*H) * var;
                */
                TooN::Vector<2> K = var.T()[0] / (obsVar + var(0,0));   //K is first col = first row of var.
                state = state + K * (obs - state[0]);
                TooN::Matrix<2> tmp = TooN::Identity;
                tmp(0,0) -= K[0];
                tmp(1,0) -= K[1];
                var = tmp * var;

        }


        inline void observeSpeed(double obs, double obsVar)
        {
                /* MATLAB:
                H = [0 1];
                K = (uncertainty * H') / ((H * uncertainty * H') + obsVar);
                state = state + K * (observation - H*state);
                uncertainty = (eye(2)-K*H) * uncertainty;
                */
                TooN::Vector<2> K = var.T()[1] / (obsVar + var(1,1));   //K is second col = second row of var.
                state = state + K * (obs - state[1]);
                TooN::Matrix<2> tmp = TooN::Identity;
                tmp(0,1) -= K[0];
                tmp(1,1) -= K[1];
                var = tmp * var;
        }


        // predict
        // calculates prediction variance matrix based on gaussian acceleration as error.
        inline void predict(double ms, double accelerationVar, TooN::Vector<2> controlGains = TooN::makeVector(0,0), double coVarFac = 1, double speedVarFac = 1)
        {
                /* MATLAB:
                G = [1 ms/1000; 0 1];
                E = [((ms/1000)^2)/2; ms/1000]; % assume normal distributed constant ACCELERATION.
                state = G*state;
                var = G * var * G' + accelerationVarPerS*(E*E');
                */

                ms /= 1000;

                TooN::Matrix<2,2> G = TooN::Identity;
                G(0,1) = ms;

                state = G * state + controlGains;
                var  = G * var * G.T();
                var(0,0) += accelerationVar * 0.25 * ms*ms*ms*ms;
                var(1,0) += coVarFac * accelerationVar * 0.5 * ms*ms*ms;
                var(0,1) += coVarFac * accelerationVar * 0.5 * ms*ms*ms;
                var(1,1) += speedVarFac * accelerationVar * 1 * ms*ms;
        }

        // predict
        // calculates prediction using the given uncertainty matrix
        // vars is var(0) var(1) covar(0,1)
        inline void predict(double ms, TooN::Vector<3> vars, TooN::Vector<2> controlGains = TooN::makeVector(0,0))
        {
                /* MATLAB:
                G = [1 ms/1000; 0 1];
                E = [((ms/1000)^2)/2; ms/1000]; % assume normal distributed constant ACCELERATION.
                state = G*state;
                var = G * var * G' + accelerationVarPerS*(E*E');
                */

                ms /= 1000;

                TooN::Matrix<2,2> G = TooN::Identity;
                G(0,1) = ms;

                state = G * state + controlGains;
                var  = G * var * G.T();
                var(0,0) += vars[0];
                var(1,0) += vars[2];
                var(0,1) += vars[2];
                var(1,1) += vars[1];
        }
};


// KalmanFilter with only one component (pose, is observed directly)
class PFilter
{
public:
        double state;
        double var;

        inline PFilter() : state(0), var(1e10) {};
        inline PFilter(double initState) : state(initState), var(0) {};

        inline void predict(double ms, double speedVar, double controlGains = 0)
        {
                /* MATLAB:
                state = state;
                var = var + speedVar*((ms/1000)^2);
                */
                state += controlGains;
                var += speedVar * ms * ms / 1000000;
        }

        inline void observe(double obs, double obsVar)
        {
                /* MATLAB:
                obs_w = var / (var + obsVar);
                state = state * (1-obs_w) + obs * obs_w;
                var = var*obsVar / (var + obsVar);
                */
                double w = var / (var + obsVar);
                state = (1-w) * state + w * obs;
                var = var * obsVar / (var + obsVar);
        }
};


class DroneKalmanFilter
{
private:
        // filter
        PVFilter x;
        PVFilter y;
        PVFilter z;
        PFilter roll;
        PFilter pitch;
        PVFilter yaw;


        // relation parameters (ptam to imu scale / offset)
        double x_offset, y_offset, z_offset;
        double xy_scale, z_scale;
        double scale_from_z;
        double scale_from_xy;
        double roll_offset, pitch_offset, yaw_offset;
        bool offsets_xyz_initialized;
        bool scale_xyz_initialized;

        // intermediate values for re-estimation of relaton parameters
        double xyz_sum_IMUxIMU;
        double xyz_sum_PTAMxPTAM;
        double xyz_sum_PTAMxIMU;
        double rp_offset_framesContributed;
        std::vector<ScaleStruct>* scalePairs;

        // parameters used for height and yaw differentiation
        double last_z_IMU;
        long last_yaw_droneTime;
        long last_z_droneTime;
        long last_z_packageID;

        // contains the last ptam-pose added (raw ptam-data).
        TooN::Vector<3> last_PTAM_pose;

        // contains the pose directly after the last ptam-fuse.
        TooN::Vector<3> last_fused_pose;
        bool lastPosesValid;


        // statistics parameters
        int numGoodIMUObservations;
        int numGoodPTAMObservations;

        // parameters used for adding / timing
        long lastIMU_XYZ_dronetime;
        long lastIMU_RPY_dronetime;
        long lastIMU_dronetime;
        int lastIMU_XYZ_ID;
        int lastIMU_RPY_ID;
        double lastPredictedRoll;
        double lastPredictedPitch;
        double initialScaleSet;

        // internal add functions
        void predictInternal(geometry_msgs::Twist activeControlInfo, int timeSpanMicros, bool useControlGains = true);
        void observeIMU_XYZ(const ardrone_autonomy::Navdata* nav);
        void observeIMU_RPY(const ardrone_autonomy::Navdata* nav);
        void observePTAM(TooN::Vector<6> pose);


        // internal sync functions. called on ptam-add.
        void sync_xyz(double x_global, double y_global, double z_global);
        void sync_rpy(double roll_global, double pitch_global, double yaw_global);

        // ms state of filter
        int predictedUpToTotal;
        long predictedUpToDroneTime;

        // logging
        double lastdYaw, lastdZ;
        double baselineZ_IMU;
        double baselineZ_Filter;
        double last_z_heightDiff;
        double baselineY_IMU;
        bool baselinesYValid;
        int timestampYawBaselineFrom;
        double lastVXGain;
        double lastVYGain;



        EstimationNode* node;
public:
        DroneKalmanFilter(EstimationNode* node);
        ~DroneKalmanFilter(void);
        void release();

        static int delayRPY;    // assumed 0 (fastest data).
        static int delayXYZ;    // assumed 70 (gets here 70ms later than rpy)
        static int delayVideo;  // assumed 120 (gets here 120ms later than rpy)
        static int delayControl;        // assumed 120 (gets here 120ms later than rpy)

        static const int base_delayXYZ;
        static const int base_delayVideo;
        static const int base_delayControl;

        std::deque<ardrone_autonomy::Navdata>* navdataQueue;    // new navdata messages
        std::deque<geometry_msgs::TwistStamped>* velQueue;              // new velocity messages
        static pthread_mutex_t filter_CS;


        int predictdUpToTimestamp;
        int scalePairsIn, scalePairsOut;



        // resets everything to zero.
        void reset();


        // resets everything to do with PTAM to zero (call if tracking continues, but PTAM tracking is reset)
        void clearPTAM();

        // predicts up to a specified time in ms, using all available data.
        // if consume=false, does not delete anything from queues.
        void predictUpTo(int timestamp, bool consume = true, bool useControlGains = true);

        void setPing(unsigned int navPing, unsigned int vidPing);

        // gets current pose and variances (up to where predictUpTo has been called)
        TooN::Vector<6> getCurrentPose();
        uga_tum_ardrone::filter_state getCurrentPoseSpeed();
        TooN::Vector<10> getCurrentPoseSpeedAsVec();
        TooN::Vector<10> getCurrentPoseSpeedVariances();
        TooN::Vector<6> getCurrentPoseVariances();
        TooN::Vector<6> getCurrentOffsets();
        TooN::Vector<3> getCurrentScales();
        TooN::Vector<3> getCurrentScalesForLog();
        float getScaleAccuracy();
        void setCurrentScales(TooN::Vector<3> scales);

        // adds a PTAM observation. automatically predicts up to timestamp.
        void updateScaleXYZ(TooN::Vector<3> ptamDiff, TooN::Vector<3> imuDiff, TooN::Vector<3> OrgPtamPose);


        // does not actually change the state of the filter.
        // makes a compy of it, flushes all queued navdata into it, then predicts up to timestamp.
        // relatively costly (!)

        // transforms a PTAM observation.
        // translates from front to center, scales and adds offsets.
        TooN::Vector<3> transformPTAMObservation(double x,double y,double z);
        TooN::Vector<3> transformPTAMObservation(double x,double y,double z, double yaw);
        TooN::Vector<6> transformPTAMObservation(TooN::Vector<6> obs);
        TooN::Vector<6> backTransformPTAMObservation(TooN::Vector<6> obs);

        inline int getNumGoodPTAMObservations() {return numGoodPTAMObservations;}

        // when scaling factors are updated, exacly one point stays the same.
        // if useScalingFixpoint, this point is the current PTAM pose, otherwise it is sclingFixpoint (which is a PTAM-coordinate(!))
        TooN::Vector<3> scalingFixpoint;        // in PTAM's system (!)
        bool useScalingFixpoint;

        //
        void flushScalePairs();

        // locking
        bool allSyncLocked;
        bool useControl;
        bool useNavdata;
        bool usePTAM;

        // motion model parameters
        float c1;
        float c2;
        float c3;
        float c4;
        float c5;
        float c6;
        float c7;
        float c8;

    // landed drift handling
    float yawDriftThreshold;
    float zDriftThreshold;


        bool handleCommand(std::string s);

        // new ROS interface functions
        void addPTAMObservation(TooN::Vector<6> trans, int time);
        void addFakePTAMObservation(int time);
        uga_tum_ardrone::filter_state getPoseAt(ros::Time t, bool useControlGains = true);
        TooN::Vector<10> getPoseAtAsVec(int timestamp, bool useControlGains = true);

};
#endif /* __DRONEKALMANFILTER_H */