Retrodict filter to previous time step
retrodict function performs retrodiction, predicting
the state estimate and covariance backward to the time at which an out-of-sequence measurement
(OOSM) was taken. To use this function, specify the
property of the filter as a positive integer. After using this function, use the
function to update the current state estimates using the OOSM.
also returns the status of the retrodiction
retrodictStatus] = retrodict(___)
true for success and
false for failure. The
retrodiction process can fail if the length of the state history stored in the filter
(specified by the
MaxNumOOSMSteps property of the filter) does not
cover the request time specified by the
Improve Filter Estimates Using Retrodiction
Generate a truth trajectory using the 3-D constant velocity model.
rng(2021) % For repeatable results initialState = [1; 0.4; 2; 0.3; 1; -0.2]; % [x; vx; y; vy; z; vz] dt = 1; % Time step steps = 10; sigmaQ = 0.2; % Standard deviation for process noise states = NaN(6,steps); states(:,1) = initialState; for ii = 2:steps w = sigmaQ*randn(3,1); states(:,ii) = constvel(states(:,ii-1),w,dt); end
Generate position measurements from the truths.
positionSelector = [1 0 0 0 0 0; 0 0 1 0 0 0; 0 0 0 0 1 0]; sigmaR = 0.2; % Standard deviation for measurement noise positions = positionSelector*states; measures = positions + sigmaR*randn(3,steps);
Show the truths and measurements in an x-y plot.
figure plot(positions(1,:),positions(2,:),"ro","DisplayName","Truths"); hold on; plot(measures(1,:),measures(2,:),"bx","DisplayName","Measures"); xlabel("x (m)") ylabel("y (m)") legend("Location","northwest")
Assume that, at the ninth step, the measurement is delayed and therefore unavailable.
delayedMeasure = measures(:,9); measures(:,9) = NaN;
Construct an extended Kalman filter (EKF) based on the constant velocity model.
estimates = NaN(6,steps); covariances = NaN(6,6,steps); estimates(:,1) = positionSelector'*measures(:,1); covariances(:,:,1) = 1*eye(6); filter = trackingEKF(@constvel,@cvmeas,... "State",estimates(:,1),... "StateCovariance",covariances(:,:,1),... "ProcessNoise",eye(6),... "MeasurementNoise",sigmaR^2*eye(3),... "MaxNumOOSMSteps",3);
Step through the EKF with the measurements.
for ii = 2:steps predict(filter); if ~any(isnan(measures(:,ii))) % Skip if unavailable correct(filter,measures(:,ii)); end estimates(:,ii) = filter.State; covariances(:,:,ii) = filter.StateCovariance; end
Show the estimated results.
Retrodict to the ninth step, and correct the current estimates by using the out-of-sequence measurements at the ninth step.
[retroState,retroCov] = retrodict(filter,-1); [retroCorrState,retroCorrCov] = retroCorrect(filter,delayedMeasure);
Plot the retrodicted state for the ninth step.
plot([retroState(1);retroCorrState(1)],... [retroState(3),retroCorrState(3)],... "kd","DisplayName","Retrodicted")
You can use the determinant of the final state covariance to see the improvements made by retrodiction. A smaller covariance determinant indicates improved state estimates.
detWithoutRetrodiciton = det(covariances(:,:,end))
detWithoutRetrodiciton = 3.2694e-04
detWithRetrodiciton = det(retroCorrCov)
detWithRetrodiciton = 2.6063e-04
dt — Retrodiction time
Retrodiction time, in seconds, specified as a nonpositive integer. Specify retrodiction time as the time difference between the time at which the OOSM was taken and the current time.
retroState — Retrodicted state
M-by-1 real-valued vector
Retrodicted state, returned as an M-by-1 real-valued vector, where M is the size of the filter state.
retroCov — Retrodicted state covariance
M-by-M real-valued positive-definite
Retrodicted state covariance, returned as an M-by-M real-valued positive-definite matrix.
retrodictStatus — Retrodiction status
Retrodiction status, returned as
true indicating success and
false indicating failure.
Retrodiction and Retro-Correction
Assume the current time step of the filter is k. At time k, the posteriori state and state covariance of the filter are x(k|k) and P(k|k), respectively. An out-of-sequence measurement (OOSM) taken at time β now arrives at time k. Find l such that β is a time step between these two consecutive time steps:
where l is a positive integer and l < k.
In the retrodiction step, the current state and state covariance at time k are predicted back to the time of the OOSM. You can obtain the retrodicted state by propagating the state transition function backward in time. For a linear state transition function, the retrodicted state is expressed as:
where F(β,k) is the backward state transition matrix from time step k to time step β. The retrodicted covariance is obtained as:
where Q(k,β) is the covariance matrix for the process noise and,
Here, P(k|k-l) is the priori state covariance at time k, predicted from the covariance information at time k–l, and
In the second step, retro-correction, the current state and state covariance are corrected using the OOSM. The corrected state is obtained as:
where z(β) is the OOSM at time β and W(k,β), the filter gain, is expressed as:
You can obtain the equivalent measurement at time β based on the state estimate at the time k, z(β|k), as
In these expressions, R(β) is the measurement covariance matrix for the OOSM and:
where H(β) is the measurement Jacobian matrix. The corrected covariance is obtained as:
 Bar-Shalom, Y., Huimin Chen, and M. Mallick. “One-Step Solution for the Multistep out-of-Sequence-Measurement Problem in Tracking.” IEEE Transactions on Aerospace and Electronic Systems 40, no. 1 (January 2004): 27–37. https://doi.org/10.1109/TAES.2004.1292140.
C/C++ Code Generation
Generate C and C++ code using MATLAB® Coder™.
In code generation, after calling the filter, you cannot change its