DroneKalmanFilter.cpp

Go to the documentation of this file.

#include "DroneKalmanFilter.h"
#include "EstimationNode.h"

// constants (variances assumed to be present)
const double varSpeedObservation_xy = 2*2;
const double varPoseObservation_xy = 0.2*0.2;
const double varAccelerationError_xy = 8*8;

const double varPoseObservation_z_PTAM = 0.08*0.08;
const double varPoseObservation_z_IMU = 0.25*0.25;
const double varPoseObservation_z_IMU_NO_PTAM = 0.1*0.1;
const double varAccelerationError_z = 1*1;

const double varPoseObservation_rp_PTAM = 3*3;
const double varPoseObservation_rp_IMU = 1*1;
const double varSpeedError_rp = 360*360 * 16;   // increased because prediction based on control command is damn inaccurate.

const double varSpeedObservation_yaw = 5*5;
const double varPoseObservation_yaw = 3*3;
const double varAccelerationError_yaw = 360*360;
        

// constants (assumed delays in ms).
// default ping values: nav=25, vid=50
int DroneKalmanFilter::delayRPY = 0;            // always zero
int DroneKalmanFilter::delayXYZ = 40;           // base_delayXYZ, but at most delayVideo
int DroneKalmanFilter::delayVideo = 75;         // base_delayVideo + delayVid - delayNav
int DroneKalmanFilter::delayControl = 100;      // base_delayControl + 2*delayNav

const int DroneKalmanFilter::base_delayXYZ = 40;                // t_xyz - t_rpy = base_delayXYZ
const int DroneKalmanFilter::base_delayVideo = 50;              // t_cam - t_rpy = base_delayVideo + delayVid - delayNav
const int DroneKalmanFilter::base_delayControl = 50;    // t_control + t_rpy - 2*delayNav

// constants (some more parameters)
const double max_z_speed = 2.5; // maximum height speed tolerated (in m/s, everything else is considered to be due to change in floor-height).
const double scaleUpdate_min_xyDist = 0.5*0.5*0.5*0.5;
const double scaleUpdate_min_zDist = 0.1*0.1;

using namespace std;

// transform degree-angle to satisfy min <= angle < sup
double angleFromTo(double angle, double min, double sup)
{
        while(angle < min) angle += 360;
        while(angle >=  sup) angle -= 360;
        return angle;
}





pthread_mutex_t DroneKalmanFilter::filter_CS = PTHREAD_MUTEX_INITIALIZER;

DroneKalmanFilter::DroneKalmanFilter(EstimationNode* n)
{
        scalePairs = new std::vector<ScaleStruct>();
        navdataQueue = new std::deque<ardrone_autonomy::Navdata>();
        velQueue = new std::deque<geometry_msgs::TwistStamped>();

        useScalingFixpoint = false;

        this->node = n;

        pthread_mutex_lock( &filter_CS );
        reset();
        pthread_mutex_unlock( &filter_CS );





        c1 = 0.58;
        c2 = 17.8;
        c3 = 10;
        c4 = 35;
        c5 = 10;
        c6 = 25;
        c7 = 1.4;
        c8 = 1.0;

}


DroneKalmanFilter::~DroneKalmanFilter(void)
{
        // dont delete nothing here, as this is also called for shallow copy.
}

void DroneKalmanFilter::release()
{
        delete scalePairs;
        delete navdataQueue;
        delete velQueue;
}


void DroneKalmanFilter::setPing(unsigned int navPing, unsigned int vidPing)
{
        // add a constant of 20ms // 40ms to accound for delay due to ros.
        // very, very rough approximation.
        navPing += 20;
        vidPing += 40;

        int new_delayXYZ = base_delayXYZ;
        int new_delayVideo = base_delayVideo + vidPing/(int)2 - navPing/(int)2;
        int new_delayControl = base_delayControl + navPing;
        
        delayXYZ = std::min(500,std::max(40,std::min(new_delayVideo,new_delayXYZ)));
        delayVideo = std::min(500,std::max(40,new_delayVideo));
        delayControl = std::min(200,std::max(50,new_delayControl));

        std::cout << "new delasXYZ: " << delayXYZ << ", delayVideo: " << delayVideo << ", delayControl: " << delayControl << std::endl;
}

void DroneKalmanFilter::reset()
{
        // init filter with pose 0 (var 0) and speed 0 (var large).
        x = y = z = yaw = PVFilter(0);
        roll = pitch = PFilter(0);
        lastIMU_XYZ_ID = lastIMU_RPY_ID = -1;
        predictedUpToDroneTime = 0;
        last_z_heightDiff = 0;
        scalePairsIn = scalePairsOut = 0;

        // set statistic parameters to zero
        numGoodIMUObservations = 0;

        // set last times to 0, indicating that there was no prev. package.
        lastIMU_XYZ_dronetime = lastIMU_RPY_dronetime = 0;
        lastIMU_dronetime = 0;

        // clear PTAM
        clearPTAM();

        // clear IMU-queus
        navdataQueue->clear();
        velQueue->clear();
        
        predictdUpToTimestamp = getMS(ros::Time::now());
        predictedUpToTotal = -1;

        baselineZ_Filter = baselineZ_IMU = -999999;
        baselinesYValid = false;


        node->publishCommand("u l EKF has been reset to zero.");
}

void DroneKalmanFilter::clearPTAM()
{
        // offsets and scales are not initialized.
        offsets_xyz_initialized = scale_xyz_initialized = false;
        xy_scale = z_scale = scale_from_xy = scale_from_z = 1;
        roll_offset = pitch_offset = yaw_offset = x_offset = y_offset = z_offset = 0;

        xyz_sum_IMUxIMU = 0.1;
        xyz_sum_PTAMxPTAM = 0.1;
        xyz_sum_PTAMxIMU = 0.1;
        scalePairs->clear();
        scalePairsIn = 1;
        scalePairsOut = 0;

        rp_offset_framesContributed = 0;

        // set statistic parameters to zero
        numGoodPTAMObservations = 0;

        // indicates that last PTAM frame was not valid
        lastPosesValid = false;
}



// this function does the actual work, predicting one timestep ahead.
void DroneKalmanFilter::predictInternal(geometry_msgs::Twist activeControlInfo, int timeSpanMicros, bool useControlGains)
{
        if(timeSpanMicros <= 0) return;
        
        useControlGains = useControlGains && this->useControl;

        bool controlValid = !(activeControlInfo.linear.z > 1.01 || activeControlInfo.linear.z < -1.01 ||
                        activeControlInfo.linear.x > 1.01 || activeControlInfo.linear.x < -1.01 ||
                        activeControlInfo.linear.y > 1.01 || activeControlInfo.linear.y < -1.01 ||
                        activeControlInfo.angular.z > 1.01 || activeControlInfo.angular.z < -1.01);


        double tsMillis = timeSpanMicros / 1000.0;      // in milliseconds
        double tsSeconds = tsMillis / 1000.0;   // in seconds


        // predict roll, pitch, yaw
        float rollControlGain = tsSeconds*c3*(c4 * max(-0.5, min(0.5, (double)activeControlInfo.linear.y)) - roll.state);
        float pitchControlGain = tsSeconds*c3*(c4 * max(-0.5, min(0.5, (double)activeControlInfo.linear.x)) - pitch.state);
        float yawSpeedControlGain = tsSeconds*c5*(c6 * activeControlInfo.angular.z - yaw.state[1]);     // at adaption to ros, this has to be reverted for some reason....



        double yawRad = yaw.state[0] * 3.14159268 / 180;
        double rollRad = roll.state * 3.14159268 / 180;
        double pitchRad = pitch.state * 3.14159268 / 180;
        double forceX = cos(yawRad) * sin(rollRad) * cos(pitchRad) - sin(yawRad) * sin(pitchRad);
        double forceY = - sin(yawRad) * sin(rollRad) * cos(pitchRad) - cos(yawRad) * sin(pitchRad);
        

        double vx_gain = tsSeconds * c1 * (c2*forceX - x.state[1]);
        double vy_gain = tsSeconds * c1 * (c2*forceY - y.state[1]);
        double vz_gain = tsSeconds * c7 * (c8*activeControlInfo.linear.z*(activeControlInfo.linear.z < 0 ? 2 : 1) - z.state[1]);

        lastVXGain = vx_gain;
        lastVYGain = vy_gain;
        lastPredictedRoll = roll.state;
        lastPredictedPitch = pitch.state;

        if(!useControlGains || !controlValid)
        {
                vx_gain = vy_gain = vz_gain = 0;
                rollControlGain = pitchControlGain = yawSpeedControlGain = 0;
        }


        yaw.state[0] =  angleFromTo(yaw.state[0],-180,180);
        roll.predict(tsMillis,varSpeedError_rp, rollControlGain);
        pitch.predict(tsMillis,varSpeedError_rp, pitchControlGain);
        yaw.predict(tsMillis,varAccelerationError_yaw,TooN::makeVector(tsSeconds*yawSpeedControlGain/2,yawSpeedControlGain),1,5*5);
        yaw.state[0] =  angleFromTo(yaw.state[0],-180,180);

        x.predict(tsMillis,varAccelerationError_xy,TooN::makeVector(tsSeconds*vx_gain/2,vx_gain),0.0001);
        y.predict(tsMillis,varAccelerationError_xy,TooN::makeVector(tsSeconds*vy_gain/2,vy_gain),0.0001);
        z.predict(tsMillis,TooN::makeVector(tsSeconds*tsSeconds*tsSeconds*tsSeconds, 9*tsSeconds,tsSeconds*tsSeconds*tsSeconds*3),
                TooN::makeVector(tsSeconds*vz_gain/2,vz_gain));
}



void DroneKalmanFilter::observeIMU_XYZ(const ardrone_autonomy::Navdata* nav)
{

        // --------------- now: update  ---------------------------
        // transform to global CS, using current yaw
        double yawRad = yaw.state[0] * 3.14159268 / 180;
        double vx_global = (sin(yawRad) * nav->vx + cos(yawRad) * nav->vy) / 1000.0;
        double vy_global = (cos(yawRad) * nav->vx - sin(yawRad) * nav->vy) / 1000.0;

        if(vx_global > 10)
                cout << "err";

        // update x,y:
        // if PTAM isGood, assume "normal" accuracy. if not, assume very accurate speeds
        // to simulate "integrating up".
        double lastX = x.state[0];
        double lastY = y.state[0];
        if(lastPosesValid)
        {
                x.observeSpeed(vx_global,varSpeedObservation_xy*50);
                y.observeSpeed(vy_global,varSpeedObservation_xy*50);
        }
        else
        {
                x.observeSpeed(vx_global,varSpeedObservation_xy);
                y.observeSpeed(vy_global,varSpeedObservation_xy);
        }



        if(abs(lastX-x.state[0]) > 0.2 && lastX != 0)
        {
                // this happens if there was no navdata for a long time -> EKF variances got big -> new update leads to large jump in pos.
                ROS_WARN("detected large x jump. removing. should not happen usually (only e.g. if no navdata for a long time, or agressive re-scaling)");
                x.state[0] = lastX;
        }
        if(abs(lastY-y.state[0]) > 0.2 && lastY != 0)
        {
                y.state[0] = lastY;
                ROS_WARN("detected large y jump. removing. should not happen usually (only e.g. if no navdata for a long time, or agressive re-scaling)");
        }

        // height is a bit more complicated....
        // only update every 8 packages, or if changed.
        if(last_z_IMU != nav->altd || nav->header.seq - last_z_packageID > 8)
        {
                if(baselineZ_Filter < -100000)  // only for initialization.
                {
                        baselineZ_IMU = nav->altd;
                        baselineZ_Filter = z.state[0];
                }



                if(lastPosesValid)
                {

                        double imuHeightDiff = (nav->altd - baselineZ_IMU )*0.001;      // TODO negative heights??
                        double observedHeight = baselineZ_Filter + 0.5*(imuHeightDiff + last_z_heightDiff);
                        last_z_heightDiff = imuHeightDiff;

                        baselineZ_IMU = nav->altd;
                        baselineZ_Filter = z.state[0];

                        if((abs(imuHeightDiff) < 0.150 && abs(last_z_heightDiff) < 0.150))      // jumps of more than 150mm in 40ms are ignored
                        {
                                z.observePose(observedHeight,varPoseObservation_z_IMU);
                                lastdZ = observedHeight;
                        }
                }
                else
                {
                        double imuHeightDiff = (nav->altd - baselineZ_IMU )*0.001;
                        double observedHeight = baselineZ_Filter + imuHeightDiff;

                        if(abs(imuHeightDiff) < 0.110)  // jumps of more than 150mm in 40ms are ignored
                        {
                                z.observePose(observedHeight,varPoseObservation_z_IMU_NO_PTAM);
                                lastdZ = observedHeight;
                        }
                        else    // there was a jump: dont observe anything, but set new baselines.
                        {
                                if(baselineZ_IMU == 0 || nav->altd == 0)
                                {
                                        z.observePose(observedHeight,0);
                                        z.observeSpeed(0,0);
                                }

                                baselineZ_IMU = nav->altd;
                                baselineZ_Filter = z.state[0];
                        }
                }

                last_z_IMU = nav->altd;
                last_z_packageID = nav->header.seq;
        }

}



void DroneKalmanFilter::observeIMU_RPY(const ardrone_autonomy::Navdata* nav)
{
        roll.observe(nav->rotX,varPoseObservation_rp_IMU);
        pitch.observe(nav->rotY,varPoseObservation_rp_IMU);


        if(!baselinesYValid)    // only for initialization.
        {
                baselineY_IMU = nav->rotZ;
                baselineY_Filter = yaw.state[0];
                baselinesYValid = true;
                timestampYawBaselineFrom = getMS(nav->header.stamp);
        }

        double imuYawDiff = (nav->rotZ - baselineY_IMU );
        double observedYaw = baselineY_Filter + imuYawDiff;

        yaw.state[0] =  angleFromTo(yaw.state[0],-180,180);

        if(yaw.state[0] < -90)
                observedYaw = angleFromTo(observedYaw,-360,0);
        else if(yaw.state[0] > 90)
                observedYaw = angleFromTo(observedYaw,0,360);
        else
                observedYaw = angleFromTo(observedYaw,-180,180);


        if(lastPosesValid)
        {

                baselineY_IMU = nav->rotZ;
                baselineY_Filter = yaw.state[0];
                timestampYawBaselineFrom = getMS(nav->header.stamp);


                if(abs(observedYaw - yaw.state[0]) < 10)
                {
                        yaw.observePose(observedYaw,2*2);
                        lastdYaw = observedYaw;
                }
        }
        else
                if(abs(observedYaw - yaw.state[0]) < 10)
                {
                        yaw.observePose(observedYaw,1*1);
                        lastdYaw = observedYaw;
                }
                else
                {
                        baselineY_IMU = nav->rotZ;
                        baselineY_Filter = yaw.state[0];
                        timestampYawBaselineFrom = getMS(nav->header.stamp);
                }

        last_yaw_IMU = nav->rotZ;
        yaw.state[0] =  angleFromTo(yaw.state[0],-180,180);
}



void DroneKalmanFilter::observePTAM(TooN::Vector<6> pose)
{

        // ----------------- observe xyz -----------------------------
        // sync!
        if(!allSyncLocked)
        {
                sync_xyz(pose[0], pose[1], pose[2]);
                sync_rpy(pose[3], pose[4], pose[5]);
        }

        last_PTAM_pose = TooN::makeVector(pose[0], pose[1], pose[2]);


        pose.slice<0,3>() = transformPTAMObservation(pose[0], pose[1], pose[2]);

        if(offsets_xyz_initialized)
        {
                x.observePose(pose[0],varPoseObservation_xy);
                y.observePose(pose[1],varPoseObservation_xy);
        }

        // observe z
        if(offsets_xyz_initialized)
        {
                z.observePose(pose[2], varPoseObservation_z_PTAM);
        }

        // observe!
        if(rp_offset_framesContributed > 1)
        {
                roll.observe(pose[3]+roll_offset,varPoseObservation_rp_PTAM);
                pitch.observe(pose[4]+pitch_offset,varPoseObservation_rp_PTAM);

                yaw.state[0] =  angleFromTo(yaw.state[0],-180,180);
                double observedYaw = pose[5]+yaw_offset;

                if(yaw.state[0] < -90)
                        observedYaw = angleFromTo(observedYaw,-360,0);
                else if(yaw.state[0] > 90)
                        observedYaw = angleFromTo(observedYaw,0,360);
                else
                        observedYaw = angleFromTo(observedYaw,-180,180);

                yaw.observePose(observedYaw,varPoseObservation_yaw);
                yaw.state[0] =  angleFromTo(yaw.state[0],-180,180);
        }

        last_fused_pose = TooN::makeVector(x.state[0], y.state[0], z.state[0]);
        lastPosesValid = true;
}


void DroneKalmanFilter::sync_rpy(double roll_global, double pitch_global, double yaw_global)
{
        if(allSyncLocked) return;
        // set yaw on first call
        if(rp_offset_framesContributed < 1)
                yaw_offset = yaw.state[0] - yaw_global;
        
        // update roll and pitch offset continuously as normal average.
        if(rp_offset_framesContributed < 100)
        {
                roll_offset = rp_offset_framesContributed * roll_offset + roll.state - roll_global;
                pitch_offset = rp_offset_framesContributed * pitch_offset + pitch.state - pitch_global;
        }
        rp_offset_framesContributed++;

        roll_offset /= rp_offset_framesContributed;
        pitch_offset /= rp_offset_framesContributed;
}

void DroneKalmanFilter::sync_xyz(double x_global, double y_global, double z_global)
{
        if(allSyncLocked) return;
        // ----------- offset: just take first available ---------
        if(!offsets_xyz_initialized)
        {
                x_offset = x.state[0] - x_global*xy_scale;
                y_offset = y.state[0] - y_global*xy_scale;
                z_offset = z.state[0] - z_global*z_scale;
                offsets_xyz_initialized = true;
        }


}



void DroneKalmanFilter::flushScalePairs()
{
        std::ofstream* fle = new std::ofstream();
        fle->open ("scalePairs.txt");
        for(unsigned int i=0;i<scalePairs->size();i++)
                (*fle) << (*scalePairs)[i].ptam[0] << " " <<(*scalePairs)[i].ptam[1] << " " <<(*scalePairs)[i].ptam[2] << " " <<
                        (*scalePairs)[i].imu[0] << " " << (*scalePairs)[i].imu[1] << " " << (*scalePairs)[i].imu[2] << std::endl;
        fle->flush();
        fle->close();
        delete fle;
}

void DroneKalmanFilter::updateScaleXYZ(TooN::Vector<3> ptamDiff, TooN::Vector<3> imuDiff, TooN::Vector<3> OrgPtamPose)
{
        if(allSyncLocked) return;

        ScaleStruct s = ScaleStruct(ptamDiff, imuDiff);

        // dont add samples that are way to small...
        if(s.imuNorm < 0.05 || s.ptamNorm < 0.05) return;


        // update running sums
        (*scalePairs).push_back(s);

        double xyz_scale_old = xy_scale;


        // find median.
        std::sort((*scalePairs).begin(), (*scalePairs).end());
        double median = (*scalePairs)[((*scalePairs).size()+1)/2].alphaSingleEstimate;

        // hack: if we have only few samples, median is unreliable (maybe 2 out of 3 are completely wrong.
        // so take first scale pair in this case (i.e. the initial scale)
        if((*scalePairs).size() < 5)
                median = initialScaleSet;

        // find sums and median.
        // do separately for xy and z and xyz-all and xyz-filtered
        double sumII = 0;
        double sumPP = 0;
        double sumPI = 0;
        double totSumII = 0;
        double totSumPP = 0;
        double totSumPI = 0;
        
        double sumIIxy = 0;
        double sumPPxy = 0;
        double sumPIxy = 0;
        double sumIIz = 0;
        double sumPPz = 0;
        double sumPIz = 0;

        int numIn = 0;
        int numOut = 0;
        for(unsigned int i=0;i<(*scalePairs).size();i++)
        {
                if((*scalePairs).size() < 5 || ((*scalePairs)[i].alphaSingleEstimate > median * 0.2 && (*scalePairs)[i].alphaSingleEstimate < median / 0.2))
                {
                        sumII += (*scalePairs)[i].ii;
                        sumPP += (*scalePairs)[i].pp;
                        sumPI += (*scalePairs)[i].pi;

                        sumIIxy += (*scalePairs)[i].imu[0]*(*scalePairs)[i].imu[0] + (*scalePairs)[i].imu[1]*(*scalePairs)[i].imu[1];
                        sumPPxy += (*scalePairs)[i].ptam[0]*(*scalePairs)[i].ptam[0] + (*scalePairs)[i].ptam[1]*(*scalePairs)[i].ptam[1];
                        sumPIxy += (*scalePairs)[i].ptam[0]*(*scalePairs)[i].imu[0] + (*scalePairs)[i].ptam[1]*(*scalePairs)[i].imu[1];
                
                        sumIIz += (*scalePairs)[i].imu[2]*(*scalePairs)[i].imu[2];
                        sumPPz += (*scalePairs)[i].ptam[2]*(*scalePairs)[i].ptam[2];
                        sumPIz += (*scalePairs)[i].ptam[2]*(*scalePairs)[i].imu[2];

                        numIn++;
                }
                else
                {
                        totSumII += (*scalePairs)[i].ii;
                        totSumPP += (*scalePairs)[i].pp;
                        totSumPI += (*scalePairs)[i].pi;
                        numOut++;
                }
        }
        xyz_sum_IMUxIMU = sumII;
        xyz_sum_PTAMxPTAM = sumPP;
        xyz_sum_PTAMxIMU = sumPI;

        double scale_Filtered = (*scalePairs)[0].computeEstimator(sumPP,sumII,sumPI,0.2,0.01);
        double scale_Unfiltered = (*scalePairs)[0].computeEstimator(sumPP+totSumPP,sumII+totSumII,sumPI+totSumPI,0.2,0.01);
        double scale_PTAMSmallVar = (*scalePairs)[0].computeEstimator(sumPP+totSumPP,sumII+totSumII,sumPI+totSumPI,0.00001,1);
        double scale_IMUSmallVar = (*scalePairs)[0].computeEstimator(sumPP+totSumPP,sumII+totSumII,sumPI+totSumPI,1,0.00001);

        
        double scale_Filtered_xy = (*scalePairs)[0].computeEstimator(sumPPxy,sumIIxy,sumPIxy,0.2,0.01);
        double scale_Filtered_z = (*scalePairs)[0].computeEstimator(sumPPz,sumIIz,sumPIz,0.2,0.01);


        scalePairsIn = numIn;
        scalePairsOut = numOut;

        printf("scale: in: %i; out: %i, filt: %.3f; xyz: %.1f < %.1f < %.1f; xy: %.1f < %.1f < %.1f; z: %.1f < %.1f < %.1f;\n", 
                numIn, numOut, scale_Filtered, 
                scale_PTAMSmallVar, scale_Unfiltered, scale_IMUSmallVar,
                (*scalePairs)[0].computeEstimator(sumPPxy,sumIIxy,sumPIxy,0.00001,1),
                scale_Filtered_xy,
                (*scalePairs)[0].computeEstimator(sumPPxy,sumIIxy,sumPIxy,1,0.00001),
                (*scalePairs)[0].computeEstimator(sumPPz,sumIIz,sumPIz,0.00001,1),
                scale_Filtered_z,
                (*scalePairs)[0].computeEstimator(sumPPz,sumIIz,sumPIz,1,0.00001)
                );


        if(scale_Filtered > 0.1)
                z_scale = xy_scale = scale_Filtered;
        else
                ROS_WARN("calculated scale is too small %.3f, disallowing!",scale_Filtered);


        scale_from_xy = scale_Filtered_xy;
        scale_from_z = scale_Filtered_z;
        // update offsets such that no position change occurs (X = x_global*xy_scale_old + offset = x_global*xy_scale_new + new_offset)
        if(useScalingFixpoint)
        {
                // fix at fixpoint
                x_offset += (xyz_scale_old - xy_scale)*scalingFixpoint[0];
                y_offset += (xyz_scale_old - xy_scale)*scalingFixpoint[1];
                z_offset += (xyz_scale_old - z_scale)*scalingFixpoint[2];
        }
        else
        {
                // fix at current pos.
                x_offset += (xyz_scale_old - xy_scale)*OrgPtamPose[0];
                y_offset += (xyz_scale_old - xy_scale)*OrgPtamPose[1];
                z_offset += (xyz_scale_old - z_scale)*OrgPtamPose[2];
        }
        scale_xyz_initialized = true;
}

float DroneKalmanFilter::getScaleAccuracy()
{
        return 0.5 + 0.5*std::min(1.0,std::max(0.0,xyz_sum_PTAMxIMU * xy_scale/4));     // scale-corrected PTAM x IMU
}


void DroneKalmanFilter::predictUpTo(int timestamp, bool consume, bool useControlGains)
{
        if(predictdUpToTimestamp == timestamp) return;

        //std::cout << (consume ? "per" : "tmp") << " pred @ " << this << ": " << predictdUpToTimestamp << " to " << timestamp << std::endl;



        // at this point:
        // - velQueue contains controls, timestamped with time at which they were sent.
        // - navQueue contains navdata, timestamped with time at which they were received.
        // - timestamp is the time up to which we want to predict, i.e. maybe a little bit into the feature

        // start at [predictdUpToTimestamp]. predict step-by-step observing at [currentTimestamp] the
        // - rpy timestamped with [currentTimestamp + delayRPY]
        // - xyz timestamped with [currentTimestamp + delayXYZ]
        // using
        // - control timestamped with [currentTimestamp - delayControl]

        // fast forward until first package that will be used.
        // for controlIterator, this is the last package with a stamp smaller/equal than what it should be.
        // for both others, thi is the first package with a stamp bigger than what it should be.
        // if consume, delete everything before permanently.
        std::deque<geometry_msgs::TwistStamped>::iterator controlIterator = velQueue->begin();
        while(controlIterator != velQueue->end() &&
                        controlIterator+1 != velQueue->end() &&
                        getMS((controlIterator+1)->header.stamp) <= predictdUpToTimestamp - delayControl)
                if(consume)
                {
                        velQueue->pop_front();
                        controlIterator = velQueue->begin();
                }
                else
                        controlIterator++;
        if(velQueue->size() == 0) useControlGains = false;

        // dont delete here, it will be deleted if respective rpy data is consumed.
        std::deque<ardrone_autonomy::Navdata>::iterator xyzIterator = navdataQueue->begin();
        while(xyzIterator != navdataQueue->end() &&
                        getMS(xyzIterator->header.stamp) <= predictdUpToTimestamp + delayXYZ)
                xyzIterator++;

        std::deque<ardrone_autonomy::Navdata>::iterator rpyIterator = navdataQueue->begin();
        while(rpyIterator != navdataQueue->end() &&
                        getMS(rpyIterator->header.stamp) <= predictdUpToTimestamp + delayRPY)
                if(consume)
                {
                        navdataQueue->pop_front();
                        rpyIterator = navdataQueue->begin();
                }
                else
                        rpyIterator++;

        // now, each iterator points to the first elemnent in queue that is to be integrated.
        // start predicting,
        while(true)
        {
                // predict ahead to [timestamp]
                int predictTo = timestamp;

                // but a maximum of 10ms per prediction step, to guarantee nonlinearities.
                predictTo = min(predictTo, predictdUpToTimestamp+10);

                // get three queues to the right point in time by rolling forward in them.
                // for xyz this is the first point at which its obs-time is bigger than or equal to [predictdUpToTimestamp]
                while(xyzIterator != navdataQueue->end() &&
                                getMS(xyzIterator->header.stamp) - delayXYZ < predictdUpToTimestamp)
                        xyzIterator++;
                while(rpyIterator != navdataQueue->end() &&
                                getMS(rpyIterator->header.stamp) - delayRPY < predictdUpToTimestamp)
                        rpyIterator++;
                // for control that is last message with stamp <= predictdUpToTimestamp - delayControl.
                while(controlIterator != velQueue->end() &&
                                controlIterator+1 != velQueue->end() &&
                                getMS((controlIterator+1)->header.stamp) + delayControl <= predictdUpToTimestamp)
                        controlIterator++;



                // predict not further than the point in time where the next observation needs to be added.
                if(rpyIterator != navdataQueue->end() )
                        predictTo = min(predictTo, getMS(rpyIterator->header.stamp)-delayRPY);
                if(xyzIterator != navdataQueue->end() )
                        predictTo = min(predictTo, getMS(xyzIterator->header.stamp)-delayXYZ);




                predictInternal(useControlGains ? controlIterator->twist : geometry_msgs::Twist(),
                                (predictTo - predictdUpToTimestamp)*1000,
                                useControlGains &&
                                getMS(controlIterator->header.stamp) + 200 > predictdUpToTimestamp - delayControl);                             // control max. 200ms old.

                //cout << " " << (predictTo - predictdUpToTimestamp);

                // if an observation needs to be added, it HAS to have a stamp equal to [predictTo],
                // as we just set [predictTo] to that timestamp.
                bool observedXYZ = false, observedRPY=false;
                if(rpyIterator != navdataQueue->end() && getMS(rpyIterator->header.stamp)-delayRPY == predictTo)
                {
                        if(this->useNavdata)
                                observeIMU_RPY(&(*rpyIterator));

                        observedRPY = true;
                        //cout << "a";
                }
                if(xyzIterator != navdataQueue->end() && getMS(xyzIterator->header.stamp)-delayXYZ == predictTo)
                {
                        if(this->useNavdata)
                                observeIMU_XYZ(&(*xyzIterator));

                        observedXYZ = true;
                        //cout << "p";
                }


                predictdUpToTimestamp = predictTo;

                if(consume)
                {
                        if(node->logfileFilter != NULL)
                        {
                                pthread_mutex_lock(&(node->logFilter_CS));
                                (*(node->logfileFilter)) << predictdUpToTimestamp << " " << 0 << " " << 0 << " " << 0 << " " <<
                                        0 << " " << 0 << " " << 0 << " " <<
                                        controlIterator->twist.linear.y << " " << controlIterator->twist.linear.x << " " << controlIterator->twist.linear.z << " " << controlIterator->twist.angular.z << " " <<
                                        (observedRPY ? rpyIterator->rotX : -1) << " " << (observedRPY ? rpyIterator->rotY : -1) << " " << (observedRPY ? lastdYaw : -1) << " " <<
                                        (observedXYZ ? xyzIterator->vx : -1) << " " << (observedXYZ ? xyzIterator->vy : -1) << " " << (observedXYZ ? lastdZ : -1) << " " <<
                                        x.state[0] << " " << y.state[0] << " " << z.state[0] << " " << roll.state << " " << pitch.state << " " << yaw.state[0] << " " << x.state[1] << " " << y.state[1] << " " << z.state[1] << " " << yaw.state[1] << " " <<
                                        lastVXGain << " " << lastVYGain << " " << "\n";
                                pthread_mutex_unlock(&(node->logFilter_CS));
                        }
                }

                if(observedRPY) rpyIterator++;
                if(observedXYZ) xyzIterator++;


                // if this is where we wanna get, quit.
                if(predictTo == timestamp)
                        break;
        }
        //cout << endl;
}


TooN::Vector<3> DroneKalmanFilter::transformPTAMObservation(double x,double y,double z)
{
        return transformPTAMObservation(x,y,z,yaw.state[0]);
}
TooN::Vector<3> DroneKalmanFilter::transformPTAMObservation(double x,double y,double z, double yaw)
{
        double yawRad = yaw * 3.14159268 / 180;
        x = x_offset + xy_scale*x - 0.2*sin(yawRad);
        y = y_offset + xy_scale*y - 0.2*cos(yawRad);
        z = z_offset + z_scale*z;
        return TooN::makeVector(x,y,z);
}

TooN::Vector<6> DroneKalmanFilter::transformPTAMObservation(TooN::Vector<6> obs)
{
        obs.slice<0,3>() = transformPTAMObservation(obs[0], obs[1], obs[2], obs[5]);
        
        obs[3] += roll_offset;
        obs[4] += pitch_offset;
        obs[5] += yaw_offset;

        return obs;
}
TooN::Vector<6> DroneKalmanFilter::backTransformPTAMObservation(TooN::Vector<6> obs)
{
        obs[3] -= roll_offset;
        obs[4] -= pitch_offset;
        obs[5] -= yaw_offset;

        double yawRad = obs[5] * 3.14159268 / 180;
        obs[0] = (- x_offset + obs[0] + 0.2*sin(yawRad))/xy_scale;
        obs[1] = (- y_offset + obs[1] + 0.2*cos(yawRad))/xy_scale;
        obs[2] = (- z_offset + obs[2])/z_scale;

        return obs;
}



TooN::Vector<6> DroneKalmanFilter::getCurrentPose()
{
        return TooN::makeVector(x.state[0], y.state[0], z.state[0], roll.state, pitch.state, yaw.state[0]);
}

tum_ardrone::filter_state DroneKalmanFilter::getCurrentPoseSpeed()
{
        tum_ardrone::filter_state s;
        s.x = x.state[0];
        s.y = y.state[0];
        s.z = z.state[0];
        s.yaw = yaw.state[0];
        s.dx = x.state[1];
        s.dy = y.state[1];
        s.dz = z.state[1];
        s.dyaw = yaw.state[1];
        s.roll = roll.state;
        s.pitch = pitch.state;

        if(s.roll*s.roll < 0.001) s.roll = 0;
        if(s.pitch*s.pitch < 0.001) s.pitch = 0;
        if(s.yaw*s.yaw < 0.001) s.yaw = 0;
        if(s.dx*s.dx < 0.001) s.dx = 0;
        if(s.dy*s.dy < 0.001) s.dy = 0;
        if(s.dz*s.dz < 0.001) s.dz = 0;
        if(s.x*s.x < 0.001) s.x = 0;
        if(s.y*s.y < 0.001) s.y = 0;
        if(s.z*s.z < 0.001) s.z = 0;

        return s;
}

TooN::Vector<10> DroneKalmanFilter::getCurrentPoseSpeedAsVec()
{
        return TooN::makeVector(x.state[0], y.state[0], z.state[0], roll.state, pitch.state, yaw.state[0],
                x.state[1], y.state[1], z.state[1],yaw.state[1]);
}

TooN::Vector<10> DroneKalmanFilter::getCurrentPoseSpeedVariances()
{
        return TooN::makeVector(x.var(0,0), y.var(0,0), z.var(0,0), roll.var, pitch.var, yaw.var(0,0),
                x.var(1,1), y.var(1,1), z.var(1,1),yaw.var(1,1));
}

TooN::Vector<6> DroneKalmanFilter::getCurrentPoseVariances()
{
        return TooN::makeVector(x.var(0,0), y.var(0,0), z.var(0,0), roll.var, pitch.var, yaw.var(0,0));
}
TooN::Vector<6> DroneKalmanFilter::getCurrentOffsets()
{
        TooN::Vector<6> res = TooN::makeVector(0,0,0,0,0,0);
        if(offsets_xyz_initialized)
                res.slice<0,3>() = TooN::makeVector(x_offset, y_offset, z_offset);
        if(rp_offset_framesContributed > 1)
                res.slice<3,3>() = TooN::makeVector(roll_offset, pitch_offset, yaw_offset);
        return res;
}
TooN::Vector<3> DroneKalmanFilter::getCurrentScales()
{
        return TooN::makeVector(scale_xyz_initialized ? xy_scale : 1, scale_xyz_initialized ? xy_scale : 1, scale_xyz_initialized ? z_scale : 1);
}
TooN::Vector<3> DroneKalmanFilter::getCurrentScalesForLog()
{
        return TooN::makeVector(scale_xyz_initialized ? scale_from_xy : 1, scale_xyz_initialized ? scale_from_z : 1, scale_xyz_initialized ? xy_scale : 1);
}

void DroneKalmanFilter::setCurrentScales(TooN::Vector<3> scales)
{
        if(allSyncLocked) return;
        xy_scale = scales[0];
        z_scale = scales[0];
        scale_from_xy = scale_from_z = scales[0];

        xyz_sum_IMUxIMU = 0.2 * scales[0];
        xyz_sum_PTAMxPTAM = 0.2 / scales[0];
        xyz_sum_PTAMxIMU = 0.2;

        (*scalePairs).clear();

        (*scalePairs).push_back(ScaleStruct(
                 TooN::makeVector(0.2,0.2,0.2) / sqrt(scales[0]),
                 TooN::makeVector(0.2,0.2,0.2) * sqrt(scales[0])
                ));



        scale_xyz_initialized = true;
        offsets_xyz_initialized = false;

        initialScaleSet = scales[0];
}

void DroneKalmanFilter::addPTAMObservation(TooN::Vector<6> trans, int time)
{
        if(time > predictdUpToTimestamp)
                predictUpTo(time, true,true);

        observePTAM(trans);
        numGoodPTAMObservations++;
}
void DroneKalmanFilter::addFakePTAMObservation(int time)
{
        if(time > predictdUpToTimestamp)
                predictUpTo(time, true,true);

        lastPosesValid = false;
}
tum_ardrone::filter_state DroneKalmanFilter::getPoseAt(ros::Time t, bool useControlGains)
{
        // make shallow copy
        DroneKalmanFilter scopy = DroneKalmanFilter(*this);

        // predict using this copy
        scopy.predictUpTo(getMS(t),false, useControlGains);

        // return value, and discard any changes made to scopy (deleting it)
        return scopy.getCurrentPoseSpeed();
}

TooN::Vector<10> DroneKalmanFilter::getPoseAtAsVec(int timestamp, bool useControlGains)
{
        // make shallow copy
        DroneKalmanFilter scopy = DroneKalmanFilter(*this);

        // predict using this copy
        scopy.predictUpTo(timestamp,false, useControlGains);

        // return value, and discard any changes made to scopy (deleting it)
        return scopy.getCurrentPoseSpeedAsVec();

}

bool DroneKalmanFilter::handleCommand(std::string s)
{

        return false;
}