1
$\begingroup$

I am starting off with a very simple Kalman filter for vision based pose estimation (PnP algorithm). The filter is inspired by the constant velocity model in this OpenCV tutorial, but I am ignoring roll, pitch and yaw for now and I am only estimating the XYZ pose.

As PnP is formulated as a non linear least squares problem, I have access to a covariance matrix from the solver I am using (Ceres), and I am using this matrix as an estimate of the measurement noise covariance matrix $R$ at each step. Process noise covariance $Q$ remains constant. My understanding of the filter is that if I obtain more and more 'good' measurements (with low $R$), the system covariance should recursively decrease; and even if the measurements worsen later, the posteriors should not worsen too much. So if I were to start from an area of bad measurements with covariance $P_1$, move to an area where I receive some good measurements and back to the area of bad measurements with covariance this time being $P_2$, $P_2 < P_1$.

On the other hand, in my formulation, my posterior covariance $P$ seems to be blindly following $R$: in my previous example of bad-good-bad areas, $P$ decreases and increases to almost the same extent, so I can't see the advantage of receiving good measurements reflected in the final covariance. On the other hand, if I reduce $Q$ even lower, I can see some variation in initial vs. final covariances, but the system is not trusting the measurements at all and is smoothing the poses out way too much. I am confused as to how to pick the best values of $Q$ and $R$, and mainly as to how I can write my filter in a way that it recognizes the advantage of getting good measurements.

$\endgroup$
3
$\begingroup$

First, let me see if I can summarize your issue to see if I understand correctly: 1) Start in area with bad measurements (high R), get P1 2) Move to area with good measurements (low R) 3) Move back to first area (high R), get P2 And you expect P2 < P1, correct?

Generally speaking, yes, at some level you can think P2 should be less than P1. But the devil is in the details: over time the filter will reach a steady state so given the same state, Q, and R it will converge to the same value regardless of how you got there. If are able to instantaneously jump between the bad and good measurement regions you should initially see P2

The other thing to consider is the magnitude of R relative to Q; if even your "good" measurement R is large the filter mostly just trusts the predictive model so the quality of your measurements doesn't affect much. You can try test with a large Q (or smaller R) to trust the measurements more; but that will also make the filter adjust more quickly after changing from good->bad measurement regions.

$\endgroup$
  • $\begingroup$ Yes, your summary is spot on. And thanks for the insight, that gives me something to think about: might have to play around with the noise values for a bit with my dataset. I am confused about the fact that the filter reaching steady state though: what decides this steady state? Is it more about staying in a region of constant-fidelity measurements.. or is it just the filter running for a long enough time? Shouldn't 'convergence' by definition depend on the path taken? $\endgroup$ – HighVoltage Oct 18 '17 at 19:34
  • $\begingroup$ steady state is defined by sensor reading, Q, and R (all constant). Convergence means the value of the filter as time goes to infinity given the same inputs (sensor, Q, R). You don't need to know how you got to the current state to know what the future states will be (again, if you're sitting still getting the same inputs). $\endgroup$ – ryan0270 Oct 19 '17 at 11:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.