EKF filter update. Now correctly scaling variance (well, if you could call this ever "correct") for repeated fusion of the same data quantity.

This commit is contained in:
Lorenz Meier
2014-08-17 20:49:46 +02:00
parent 60706500df
commit 171857811f
3 changed files with 104 additions and 18 deletions
@@ -96,7 +96,10 @@ extern "C" __EXPORT int ekf_att_pos_estimator_main(int argc, char *argv[]);
__EXPORT uint32_t millis();
__EXPORT uint64_t getMicros();
static uint64_t IMUmsec = 0;
static uint64_t IMUusec = 0;
static const uint64_t FILTER_INIT_DELAY = 1 * 1000 * 1000;
uint32_t millis()
@@ -104,6 +107,11 @@ uint32_t millis()
return IMUmsec;
}
uint64_t getMicros()
{
return IMUusec;
}
class FixedwingEstimator
{
public:
@@ -850,7 +858,8 @@ FixedwingEstimator::task_main()
}
_last_sensor_timestamp = _gyro.timestamp;
IMUmsec = _gyro.timestamp / 1e3f;
IMUmsec = _gyro.timestamp / 1e3;
IMUusec = _gyro.timestamp;
float deltaT = (_gyro.timestamp - _last_run) / 1e6f;
_last_run = _gyro.timestamp;
@@ -914,7 +923,8 @@ FixedwingEstimator::task_main()
// Copy gyro and accel
_last_sensor_timestamp = _sensor_combined.timestamp;
IMUmsec = _sensor_combined.timestamp / 1e3f;
IMUmsec = _sensor_combined.timestamp / 1e3;
IMUusec = _sensor_combined.timestamp;
float deltaT = (_sensor_combined.timestamp - _last_run) / 1e6f;
@@ -994,8 +1004,6 @@ FixedwingEstimator::task_main()
if (gps_updated) {
last_gps = _gps.timestamp_position;
orb_copy(ORB_ID(vehicle_gps_position), _gps_sub, &_gps);
perf_count(_perf_gps);
@@ -1008,11 +1016,17 @@ FixedwingEstimator::task_main()
_gps_start_time = hrt_absolute_time();
/* check if we had a GPS outage for a long time */
if (hrt_elapsed_time(&last_gps) > 5 * 1000 * 1000) {
float gps_elapsed = hrt_elapsed_time(&last_gps) / 1e6f;
const float pos_reset_threshold = 5.0f; // seconds
/* timeout of 5 seconds */
if (gps_elapsed > pos_reset_threshold) {
_ekf->ResetPosition();
_ekf->ResetVelocity();
_ekf->ResetStoredStates();
}
_ekf->updateDtGpsFilt(math::constrain((_gps.timestamp_position - last_gps) / 1e6f, 0.01f, pos_reset_threshold));
/* fuse GPS updates */
@@ -1044,6 +1058,8 @@ FixedwingEstimator::task_main()
newDataGps = true;
last_gps = _gps.timestamp_position;
}
}
@@ -1052,8 +1068,15 @@ FixedwingEstimator::task_main()
orb_check(_baro_sub, &baro_updated);
if (baro_updated) {
hrt_abstime baro_last = _baro.timestamp;
orb_copy(ORB_ID(sensor_baro0), _baro_sub, &_baro);
float baro_elapsed = (_baro.timestamp - baro_last) / 1e6f;
_ekf->updateDtHgtFilt(math::constrain(baro_elapsed, 0.001f, 0.1));
_ekf->baroHgt = _baro.altitude;
if (!_baro_init) {
@@ -2,6 +2,11 @@
#include <string.h>
#include <stdio.h>
#include <stdarg.h>
#include <math.h>
#ifndef M_PI_F
#define M_PI_F ((float)M_PI)
#endif
#define EKF_COVARIANCE_DIVERGED 1.0e8f
@@ -38,6 +43,7 @@ AttPosEKF::AttPosEKF() :
resetStates{},
storedStates{},
statetimeStamp{},
lastVelPosFusion(millis()),
statesAtVelTime{},
statesAtPosTime{},
statesAtHgtTime{},
@@ -59,7 +65,12 @@ AttPosEKF::AttPosEKF() :
accel(),
dVelIMU(),
dAngIMU(),
dtIMU(0),
dtIMU(0.005f),
dtIMUfilt(0.005f),
dtVelPos(0.01f),
dtVelPosFilt(0.01f),
dtHgtFilt(0.01f),
dtGpsFilt(0.1f),
fusionModeGPS(0),
innovVelPos{},
varInnovVelPos{},
@@ -260,6 +271,9 @@ void AttPosEKF::UpdateStrapdownEquationsNED()
// Constrain states (to protect against filter divergence)
ConstrainStates();
// update filtered IMU time step length
dtIMUfilt = 0.99f * dtIMUfilt + 0.01f * dtIMU;
}
void AttPosEKF::CovariancePrediction(float dt)
@@ -962,6 +976,21 @@ void AttPosEKF::CovariancePrediction(float dt)
ConstrainVariances();
}
void AttPosEKF::updateDtGpsFilt(float dt)
{
dtGpsFilt = ConstrainFloat(dt, 0.001f, 2.0f) * 0.05f + dtGpsFilt * 0.95f;
}
void AttPosEKF::updateDtHgtFilt(float dt)
{
dtHgtFilt = ConstrainFloat(dt, 0.001f, 2.0f) * 0.05f + dtHgtFilt * 0.95f;
}
void AttPosEKF::updateDtVelPosFilt(float dt)
{
dtVelPosFilt = ConstrainFloat(dt, 0.0005f, 2.0f) * 0.05f + dtVelPosFilt * 0.95f;
}
void AttPosEKF::FuseVelposNED()
{
@@ -998,6 +1027,18 @@ void AttPosEKF::FuseVelposNED()
// associated with sequential fusion
if (fuseVelData || fusePosData || fuseHgtData)
{
uint64_t tNow = getMicros();
updateDtVelPosFilt((tNow - lastVelPosFusion) / 1e6f);
lastVelPosFusion = tNow;
// scaler according to the number of repetitions of the
// same measurement in one fusion step
float gpsVarianceScaler = dtGpsFilt / dtVelPosFilt;
// scaler according to the number of repetitions of the
// same measurement in one fusion step
float hgtVarianceScaler = dtHgtFilt / dtVelPosFilt;
// set the GPS data timeout depending on whether airspeed data is present
if (useAirspeed) horizRetryTime = gpsRetryTime;
else horizRetryTime = gpsRetryTimeNoTAS;
@@ -1010,12 +1051,12 @@ void AttPosEKF::FuseVelposNED()
// Estimate the GPS Velocity, GPS horiz position and height measurement variances.
velErr = 0.2f*accNavMag; // additional error in GPS velocities caused by manoeuvring
posErr = 0.2f*accNavMag; // additional error in GPS position caused by manoeuvring
R_OBS[0] = sq(vneSigma) + sq(velErr);
R_OBS[0] = gpsVarianceScaler * sq(vneSigma) + sq(velErr);
R_OBS[1] = R_OBS[0];
R_OBS[2] = sq(vdSigma) + sq(velErr);
R_OBS[3] = sq(posNeSigma) + sq(posErr);
R_OBS[2] = gpsVarianceScaler * sq(vdSigma) + sq(velErr);
R_OBS[3] = gpsVarianceScaler * sq(posNeSigma) + sq(posErr);
R_OBS[4] = R_OBS[3];
R_OBS[5] = sq(posDSigma) + sq(posErr);
R_OBS[5] = hgtVarianceScaler * sq(posDSigma) + sq(posErr);
// calculate innovations and check GPS data validity using an innovation consistency check
if (fuseVelData)
@@ -1995,9 +2036,8 @@ void AttPosEKF::FuseOptFlow()
varInnovOptFlow[0] = 1.0f/SK_LOS[0];
innovOptFlow[0] = losPred[0] - losData[0];
// reset the observation index to 0 (we start by fusing the X
// measurement)
obsIndex = 0;
// set the observation index to 1 to fuse the y component next time round and reset the commence fusion flag
obsIndex = 1;
fuseOptFlowData = false;
}
else if (obsIndex == 1) // we are now fusing the Y measurement
@@ -2041,6 +2081,10 @@ void AttPosEKF::FuseOptFlow()
Kfusion[22] = SK_LOS[1]*(P[22][0]*SKK_LOS[10] - P[22][3]*SKK_LOS[13] + P[22][1]*SKK_LOS[11] + P[22][2]*SKK_LOS[12] - P[22][9]*SH_LOS[1]*SH_LOS[7]*SKK_LOS[14] + P[22][22]*SH_LOS[1]*SH_LOS[7]*SKK_LOS[14] + P[22][6]*SH_LOS[3]*SKK_LOS[0]*SKK_LOS[14] + P[22][4]*SH_LOS[3]*SKK_LOS[2]*SKK_LOS[14] - P[22][5]*SH_LOS[3]*SKK_LOS[1]*SKK_LOS[14]);
varInnovOptFlow[1] = 1.0f/SK_LOS[1];
innovOptFlow[1] = losPred[1] - losData[1];
// reset the observation index
obsIndex = 0;
fuseOptFlowData = false;
}
// Check the innovation for consistency and don't fuse if > 3Sigma
@@ -2102,10 +2146,9 @@ void AttPosEKF::FuseOptFlow()
P[i][j] = P[i][j] - KHP[i][j];
}
}
ForceSymmetry();
ConstrainVariances();
}
obsIndex = obsIndex + 1;
ForceSymmetry();
ConstrainVariances();
}
void AttPosEKF::zeroCols(float (&covMat)[n_states][n_states], uint8_t first, uint8_t last)
@@ -3006,9 +3049,14 @@ void AttPosEKF::ZeroVariables()
{
// Initialize on-init initialized variables
dtIMUfilt = ConstrainFloat(dtIMU, 0.001f, 0.02f);
dtVelPosFilt = ConstrainFloat(dtVelPos, 0.04f, 0.5f);
dtGpsFilt = 1.0f / 5.0f;
dtHgtFilt = 1.0f / 100.0f;
storeIndex = 0;
lastVelPosFusion = millis();
// Do the data structure init
for (unsigned i = 0; i < n_states; i++) {
for (unsigned j = 0; j < n_states; j++) {
@@ -116,6 +116,9 @@ public:
float storedStates[n_states][data_buffer_size]; // state vectors stored for the last 50 time steps
uint32_t statetimeStamp[data_buffer_size]; // time stamp for each state vector stored
// Times
uint64_t lastVelPosFusion; // the time of the last velocity fusion, in the standard time unit of the filter
float statesAtVelTime[n_states]; // States at the effective measurement time for posNE and velNED measurements
float statesAtPosTime[n_states]; // States at the effective measurement time for posNE and velNED measurements
float statesAtHgtTime[n_states]; // States at the effective measurement time for the hgtMea measurement
@@ -140,7 +143,12 @@ public:
Vector3f accel; // acceleration vector in XYZ body axes measured by the IMU (m/s^2)
Vector3f dVelIMU;
Vector3f dAngIMU;
float dtIMU; // time lapsed since the last IMU measurement or covariance update (sec)
float dtIMU; // time lapsed since the last IMU measurement or covariance update (sec), this may have significant jitter
float dtIMUfilt; // average time between IMU measurements (sec)
float dtVelPos; // time lapsed since the last position / velocity fusion (seconds), this may have significant jitter
float dtVelPosFilt; // average time between position / velocity fusion steps
float dtHgtFilt; // average time between height measurement updates
float dtGpsFilt; // average time between gps measurement updates
uint8_t fusionModeGPS; // 0 = GPS outputs 3D velocity, 1 = GPS outputs 2D velocity, 2 = GPS outputs no velocity
float innovVelPos[6]; // innovation output
float varInnovVelPos[6]; // innovation variance output
@@ -211,6 +219,9 @@ public:
unsigned storeIndex;
void updateDtGpsFilt(float dt);
void updateDtHgtFilt(float dt);
void UpdateStrapdownEquationsNED();
@@ -300,6 +311,8 @@ void InitializeDynamic(float (&initvelNED)[3], float declination);
protected:
void updateDtVelPosFilt(float dt);
bool FilterHealthy();
bool GyroOffsetsDiverged();
@@ -314,3 +327,5 @@ void AttitudeInit(float ax, float ay, float az, float mx, float my, float mz, fl
uint32_t millis();
uint64_t getMicros();