void Ekf::correct(const Measurement &measurement)
{
if (getDebug())
{
*debugStream_ << "---------------------- Ekf::correct ----------------------\n";
*debugStream_ << "Measurement is:\n";
*debugStream_ << measurement.measurement_ << "\n";
*debugStream_ << "Measurement covariance is:\n";
*debugStream_ << measurement.covariance_ << "\n";
}
// We don't want to update everything, so we need to build matrices that only update
// the measured parts of our state vector
// First, determine how many state vector values we're updating
std::vector<size_t> updateIndices;
for (size_t i = 0; i < measurement.updateVector_.size(); ++i)
{
if (measurement.updateVector_[i])
{
// Handle nan and inf values in measurements
if (std::isnan(measurement.measurement_(i)) || std::isnan(measurement.covariance_(i,i)))
{
if (getDebug())
{
*debugStream_ << "Value at index " << i << " was nan. Excluding from update.\n";
}
}
else if (std::isinf(measurement.measurement_(i)) || std::isinf(measurement.covariance_(i,i)))
{
if (getDebug())
{
*debugStream_ << "Value at index " << i << " was inf. Excluding from update.\n";
}
}
else
{
updateIndices.push_back(i);
}
}
}
if (getDebug())
{
*debugStream_ << "Update indices are:\n";
*debugStream_ << updateIndices << "\n";
}
size_t updateSize = updateIndices.size();
// Now set up the relevant matrices
Eigen::VectorXd stateSubset(updateSize); // x (in most literature)
Eigen::VectorXd measurementSubset(updateSize); // z
Eigen::MatrixXd measurementCovarianceSubset(updateSize, updateSize); // R
Eigen::MatrixXd stateToMeasurementSubset(updateSize, state_.rows()); // H
Eigen::MatrixXd kalmanGainSubset(state_.rows(), updateSize); // K
Eigen::VectorXd innovationSubset(updateSize); // z - Hx
stateSubset.setZero();
measurementSubset.setZero();
measurementCovarianceSubset.setZero();
stateToMeasurementSubset.setZero();
kalmanGainSubset.setZero();
innovationSubset.setZero();
// Now build the sub-matrices from the full-sized matrices
for (size_t i = 0; i < updateSize; ++i)
{
measurementSubset(i) = measurement.measurement_(updateIndices[i]);
stateSubset(i) = state_(updateIndices[i]);
for (size_t j = 0; j < updateSize; ++j)
{
measurementCovarianceSubset(i, j) = measurement.covariance_(updateIndices[i], updateIndices[j]);
}
// Handle negative (read: bad) covariances in the measurement. Rather
// than exclude the measurement or make up a covariance, just take
// the absolute value.
if (measurementCovarianceSubset(i, i) < 0)
{
if (getDebug())
{
*debugStream_ << "WARNING: Negative covariance for index " << i << " of measurement (value is" << measurementCovarianceSubset(i, i)
<< "). Using absolute value...\n";
}
measurementCovarianceSubset(i, i) = ::fabs(measurementCovarianceSubset(i, i));
}
// If the measurement variance for a given variable is very
// near 0 (as in e-50 or so) and the variance for that
// variable in the covariance matrix is also near zero, then
// the Kalman gain computation will blow up. Really, no
// measurement can be completely without error, so add a small
// amount in that case.
if (measurementCovarianceSubset(i, i) < 1e-6)
{
measurementCovarianceSubset(i, i) = 1e-6;
if (getDebug())
{
*debugStream_ << "WARNING: measurement had very small error covariance for index " << i << ". Adding some noise to maintain filter stability.\n";
}
}
}
// The state-to-measurement function, h, will now be a measurement_size x full_state_size
// matrix, with ones in the (i, i) locations of the values to be updated
for (size_t i = 0; i < updateSize; ++i)
{
stateToMeasurementSubset(i, updateIndices[i]) = 1;
}
if (getDebug())
{
*debugStream_ << "Current state subset is:\n";
*debugStream_ << stateSubset << "\n";
*debugStream_ << "Measurement subset is:\n";
*debugStream_ << measurementSubset << "\n";
*debugStream_ << "Measurement covariance subset is:\n";
*debugStream_ << measurementCovarianceSubset << "\n";
*debugStream_ << "State-to-measurement subset is:\n";
*debugStream_ << stateToMeasurementSubset << "\n";
}
// (1) Compute the Kalman gain: K = (PH') / (HPH' + R)
kalmanGainSubset = estimateErrorCovariance_ * stateToMeasurementSubset.transpose()
* (stateToMeasurementSubset * estimateErrorCovariance_ * stateToMeasurementSubset.transpose() + measurementCovarianceSubset).inverse();
// (2) Apply the gain to the difference between the state and measurement: x = x + K(z - Hx)
innovationSubset = (measurementSubset - stateSubset);
// Wrap angles in the innovation
for (size_t i = 0; i < updateSize; ++i)
{
if (updateIndices[i] == StateMemberRoll || updateIndices[i] == StateMemberPitch || updateIndices[i] == StateMemberYaw)
{
if (innovationSubset(i) < -pi_)
{
innovationSubset(i) += tau_;
}
else if (innovationSubset(i) > pi_)
{
innovationSubset(i) -= tau_;
}
}
}
state_ = state_ + kalmanGainSubset * innovationSubset;
// (3) Update the estimate error covariance using the Joseph form: (I - KH)P(I - KH)' + KRK'
Eigen::MatrixXd gainResidual = identity_ - kalmanGainSubset * stateToMeasurementSubset;
estimateErrorCovariance_ = gainResidual * estimateErrorCovariance_ * gainResidual.transpose() +
kalmanGainSubset * measurementCovarianceSubset * kalmanGainSubset.transpose();
// Handle wrapping of angles
wrapStateAngles();
if (getDebug())
{
*debugStream_ << "Kalman gain subset is:\n";
*debugStream_ << kalmanGainSubset << "\n";
*debugStream_ << "Innovation:\n";
*debugStream_ << innovationSubset << "\n\n";
*debugStream_ << "Corrected full state is:\n";
*debugStream_ << state_ << "\n";
*debugStream_ << "Corrected full estimate error covariance is:\n";
*debugStream_ << estimateErrorCovariance_ << "\n";
*debugStream_ << "Last measurement time is:\n";
*debugStream_ << std::setprecision(20) << lastMeasurementTime_ << "\n";
*debugStream_ << "\n---------------------- /Ekf::correct ----------------------\n";
}
}
void Ukf::correct(const Measurement &measurement)
{
FB_DEBUG("---------------------- Ukf::correct ----------------------\n" <<
"State is:\n" << state_ <<
"\nMeasurement is:\n" << measurement.measurement_ <<
"\nMeasurement covariance is:\n" << measurement.covariance_ << "\n");
// In our implementation, it may be that after we call predict once, we call correct
// several times in succession (multiple measurements with different time stamps). In
// that event, the sigma points need to be updated to reflect the current state.
if(!uncorrected_)
{
// Take the square root of a small fraction of the estimateErrorCovariance_ using LL' decomposition
weightedCovarSqrt_ = ((STATE_SIZE + lambda_) * estimateErrorCovariance_).llt().matrixL();
// Compute sigma points
// First sigma point is the current state
sigmaPoints_[0] = state_;
// Next STATE_SIZE sigma points are state + weightedCovarSqrt_[ith column]
// STATE_SIZE sigma points after that are state - weightedCovarSqrt_[ith column]
for(size_t sigmaInd = 0; sigmaInd < STATE_SIZE; ++sigmaInd)
{
sigmaPoints_[sigmaInd + 1] = state_ + weightedCovarSqrt_.col(sigmaInd);
sigmaPoints_[sigmaInd + 1 + STATE_SIZE] = state_ - weightedCovarSqrt_.col(sigmaInd);
}
}
// We don't want to update everything, so we need to build matrices that only update
// the measured parts of our state vector
// First, determine how many state vector values we're updating
std::vector<size_t> updateIndices;
for (size_t i = 0; i < measurement.updateVector_.size(); ++i)
{
if (measurement.updateVector_[i])
{
// Handle nan and inf values in measurements
if (std::isnan(measurement.measurement_(i)))
{
FB_DEBUG("Value at index " << i << " was nan. Excluding from update.\n");
}
else if (std::isinf(measurement.measurement_(i)))
{
FB_DEBUG("Value at index " << i << " was inf. Excluding from update.\n");
}
else
{
updateIndices.push_back(i);
}
}
}
FB_DEBUG("Update indices are:\n" << updateIndices << "\n");
size_t updateSize = updateIndices.size();
// Now set up the relevant matrices
Eigen::VectorXd stateSubset(updateSize); // x (in most literature)
Eigen::VectorXd measurementSubset(updateSize); // z
Eigen::MatrixXd measurementCovarianceSubset(updateSize, updateSize); // R
Eigen::MatrixXd stateToMeasurementSubset(updateSize, STATE_SIZE); // H
Eigen::MatrixXd kalmanGainSubset(STATE_SIZE, updateSize); // K
Eigen::VectorXd innovationSubset(updateSize); // z - Hx
Eigen::VectorXd predictedMeasurement(updateSize);
Eigen::VectorXd sigmaDiff(updateSize);
Eigen::MatrixXd predictedMeasCovar(updateSize, updateSize);
Eigen::MatrixXd crossCovar(STATE_SIZE, updateSize);
std::vector<Eigen::VectorXd> sigmaPointMeasurements(sigmaPoints_.size(), Eigen::VectorXd(updateSize));
stateSubset.setZero();
measurementSubset.setZero();
measurementCovarianceSubset.setZero();
stateToMeasurementSubset.setZero();
kalmanGainSubset.setZero();
innovationSubset.setZero();
predictedMeasurement.setZero();
predictedMeasCovar.setZero();
crossCovar.setZero();
// Now build the sub-matrices from the full-sized matrices
for (size_t i = 0; i < updateSize; ++i)
{
measurementSubset(i) = measurement.measurement_(updateIndices[i]);
stateSubset(i) = state_(updateIndices[i]);
for (size_t j = 0; j < updateSize; ++j)
{
measurementCovarianceSubset(i, j) = measurement.covariance_(updateIndices[i], updateIndices[j]);
}
// Handle negative (read: bad) covariances in the measurement. Rather
// than exclude the measurement or make up a covariance, just take
// the absolute value.
if (measurementCovarianceSubset(i, i) < 0)
{
FB_DEBUG("WARNING: Negative covariance for index " << i <<
" of measurement (value is" << measurementCovarianceSubset(i, i) <<
"). Using absolute value...\n");
measurementCovarianceSubset(i, i) = ::fabs(measurementCovarianceSubset(i, i));
}
// If the measurement variance for a given variable is very
// near 0 (as in e-50 or so) and the variance for that
// variable in the covariance matrix is also near zero, then
// the Kalman gain computation will blow up. Really, no
// measurement can be completely without error, so add a small
// amount in that case.
if (measurementCovarianceSubset(i, i) < 1e-9)
{
measurementCovarianceSubset(i, i) = 1e-9;
FB_DEBUG("WARNING: measurement had very small error covariance for index " <<
updateIndices[i] <<
". Adding some noise to maintain filter stability.\n");
}
}
// The state-to-measurement function, h, will now be a measurement_size x full_state_size
// matrix, with ones in the (i, i) locations of the values to be updated
for (size_t i = 0; i < updateSize; ++i)
{
stateToMeasurementSubset(i, updateIndices[i]) = 1;
}
FB_DEBUG("Current state subset is:\n" << stateSubset <<
"\nMeasurement subset is:\n" << measurementSubset <<
"\nMeasurement covariance subset is:\n" << measurementCovarianceSubset <<
"\nState-to-measurement subset is:\n" << stateToMeasurementSubset << "\n");
// (1) Generate sigma points, use them to generate a predicted measurement
for(size_t sigmaInd = 0; sigmaInd < sigmaPoints_.size(); ++sigmaInd)
{
sigmaPointMeasurements[sigmaInd] = stateToMeasurementSubset * sigmaPoints_[sigmaInd];
predictedMeasurement += stateWeights_[sigmaInd] * sigmaPointMeasurements[sigmaInd];
}
// (2) Use the sigma point measurements and predicted measurement to compute a predicted
// measurement covariance matrix P_zz and a state/measurement cross-covariance matrix P_xz.
for(size_t sigmaInd = 0; sigmaInd < sigmaPoints_.size(); ++sigmaInd)
{
sigmaDiff = sigmaPointMeasurements[sigmaInd] - predictedMeasurement;
predictedMeasCovar += covarWeights_[sigmaInd] * (sigmaDiff * sigmaDiff.transpose());
crossCovar += covarWeights_[sigmaInd] * ((sigmaPoints_[sigmaInd] - state_) * sigmaDiff.transpose());
}
// (3) Compute the Kalman gain, making sure to use the actual measurement covariance: K = P_xz * (P_zz + R)^-1
Eigen::MatrixXd invInnovCov = (predictedMeasCovar + measurementCovarianceSubset).inverse();
kalmanGainSubset = crossCovar * invInnovCov;
// (4) Apply the gain to the difference between the actual and predicted measurements: x = x + K(z - z_hat)
innovationSubset = (measurementSubset - predictedMeasurement);
// (5) Check Mahalanobis distance of innovation
if (checkMahalanobisThreshold(innovationSubset, invInnovCov, measurement.mahalanobisThresh_))
{
// Wrap angles in the innovation
for (size_t i = 0; i < updateSize; ++i)
{
if (updateIndices[i] == StateMemberRoll || updateIndices[i] == StateMemberPitch || updateIndices[i] == StateMemberYaw)
{
while (innovationSubset(i) < -PI)
{
innovationSubset(i) += TAU;
}
while (innovationSubset(i) > PI)
{
innovationSubset(i) -= TAU;
}
}
}
state_ = state_ + kalmanGainSubset * innovationSubset;
// (6) Compute the new estimate error covariance P = P - (K * P_zz * K')
estimateErrorCovariance_ = estimateErrorCovariance_.eval() - (kalmanGainSubset * predictedMeasCovar * kalmanGainSubset.transpose());
wrapStateAngles();
// Mark that we need to re-compute sigma points for successive corrections
uncorrected_ = false;
FB_DEBUG("Predicated measurement covariance is:\n" << predictedMeasCovar <<
"\nCross covariance is:\n" << crossCovar <<
"\nKalman gain subset is:\n" << kalmanGainSubset <<
"\nInnovation:\n" << innovationSubset <<
"\nCorrected full state is:\n" << state_ <<
"\nCorrected full estimate error covariance is:\n" << estimateErrorCovariance_ <<
"\n\n---------------------- /Ukf::correct ----------------------\n");
}
}
void Ekf::correct(const Measurement &measurement)
{
FB_DEBUG("---------------------- Ekf::correct ----------------------\n" <<
"State is:\n" << state_ << "\n"
"Topic is:\n" << measurement.topicName_ << "\n"
"Measurement is:\n" << measurement.measurement_ << "\n"
"Measurement topic name is:\n" << measurement.topicName_ << "\n\n"
"Measurement covariance is:\n" << measurement.covariance_ << "\n");
// We don't want to update everything, so we need to build matrices that only update
// the measured parts of our state vector. Throughout prediction and correction, we
// attempt to maximize efficiency in Eigen.
// First, determine how many state vector values we're updating
std::vector<size_t> updateIndices;
for (size_t i = 0; i < measurement.updateVector_.size(); ++i)
{
if (measurement.updateVector_[i])
{
// Handle nan and inf values in measurements
if (std::isnan(measurement.measurement_(i)))
{
FB_DEBUG("Value at index " << i << " was nan. Excluding from update.\n");
}
else if (std::isinf(measurement.measurement_(i)))
{
FB_DEBUG("Value at index " << i << " was inf. Excluding from update.\n");
}
else
{
updateIndices.push_back(i);
}
}
}
FB_DEBUG("Update indices are:\n" << updateIndices << "\n");
size_t updateSize = updateIndices.size();
// Now set up the relevant matrices
Eigen::VectorXd stateSubset(updateSize); // x (in most literature)
Eigen::VectorXd measurementSubset(updateSize); // z
Eigen::MatrixXd measurementCovarianceSubset(updateSize, updateSize); // R
Eigen::MatrixXd stateToMeasurementSubset(updateSize, state_.rows()); // H
Eigen::MatrixXd kalmanGainSubset(state_.rows(), updateSize); // K
Eigen::VectorXd innovationSubset(updateSize); // z - Hx
stateSubset.setZero();
measurementSubset.setZero();
measurementCovarianceSubset.setZero();
stateToMeasurementSubset.setZero();
kalmanGainSubset.setZero();
innovationSubset.setZero();
// Now build the sub-matrices from the full-sized matrices
for (size_t i = 0; i < updateSize; ++i)
{
measurementSubset(i) = measurement.measurement_(updateIndices[i]);
stateSubset(i) = state_(updateIndices[i]);
for (size_t j = 0; j < updateSize; ++j)
{
measurementCovarianceSubset(i, j) = measurement.covariance_(updateIndices[i], updateIndices[j]);
}
// Handle negative (read: bad) covariances in the measurement. Rather
// than exclude the measurement or make up a covariance, just take
// the absolute value.
if (measurementCovarianceSubset(i, i) < 0)
{
FB_DEBUG("WARNING: Negative covariance for index " << i <<
" of measurement (value is" << measurementCovarianceSubset(i, i) <<
"). Using absolute value...\n");
measurementCovarianceSubset(i, i) = ::fabs(measurementCovarianceSubset(i, i));
}
// If the measurement variance for a given variable is very
// near 0 (as in e-50 or so) and the variance for that
// variable in the covariance matrix is also near zero, then
// the Kalman gain computation will blow up. Really, no
// measurement can be completely without error, so add a small
// amount in that case.
if (measurementCovarianceSubset(i, i) < 1e-9)
{
FB_DEBUG("WARNING: measurement had very small error covariance for index " << updateIndices[i] <<
". Adding some noise to maintain filter stability.\n");
measurementCovarianceSubset(i, i) = 1e-9;
}
}
// The state-to-measurement function, h, will now be a measurement_size x full_state_size
// matrix, with ones in the (i, i) locations of the values to be updated
for (size_t i = 0; i < updateSize; ++i)
{
stateToMeasurementSubset(i, updateIndices[i]) = 1;
}
FB_DEBUG("Current state subset is:\n" << stateSubset <<
"\nMeasurement subset is:\n" << measurementSubset <<
"\nMeasurement covariance subset is:\n" << measurementCovarianceSubset <<
"\nState-to-measurement subset is:\n" << stateToMeasurementSubset << "\n");
// (1) Compute the Kalman gain: K = (PH') / (HPH' + R)
Eigen::MatrixXd pht = estimateErrorCovariance_ * stateToMeasurementSubset.transpose();
Eigen::MatrixXd hphrInv = (stateToMeasurementSubset * pht + measurementCovarianceSubset).inverse();
kalmanGainSubset.noalias() = pht * hphrInv;
innovationSubset = (measurementSubset - stateSubset);
// Wrap angles in the innovation
for (size_t i = 0; i < updateSize; ++i)
{
if (updateIndices[i] == StateMemberRoll ||
updateIndices[i] == StateMemberPitch ||
updateIndices[i] == StateMemberYaw)
{
while (innovationSubset(i) < -PI)
{
innovationSubset(i) += TAU;
}
while (innovationSubset(i) > PI)
{
innovationSubset(i) -= TAU;
}
}
}
// (2) Check Mahalanobis distance between mapped measurement and state.
if (checkMahalanobisThreshold(innovationSubset, hphrInv, measurement.mahalanobisThresh_))
{
// (3) Apply the gain to the difference between the state and measurement: x = x + K(z - Hx)
state_.noalias() += kalmanGainSubset * innovationSubset;
// (4) Update the estimate error covariance using the Joseph form: (I - KH)P(I - KH)' + KRK'
Eigen::MatrixXd gainResidual = identity_;
gainResidual.noalias() -= kalmanGainSubset * stateToMeasurementSubset;
estimateErrorCovariance_ = gainResidual * estimateErrorCovariance_ * gainResidual.transpose();
estimateErrorCovariance_.noalias() += kalmanGainSubset *
measurementCovarianceSubset *
kalmanGainSubset.transpose();
// Handle wrapping of angles
wrapStateAngles();
FB_DEBUG("Kalman gain subset is:\n" << kalmanGainSubset <<
"\nInnovation is:\n" << innovationSubset <<
"\nCorrected full state is:\n" << state_ <<
"\nCorrected full estimate error covariance is:\n" << estimateErrorCovariance_ <<
"\n\n---------------------- /Ekf::correct ----------------------\n");
}
}
