February 16, 2015

Troubleshooting bad Kalman filter results. Results project correctly, but appear behind camera

- F (transition matrix) - seems OK
- Q (system noise) - seems OK (maybe high, could increase time-scale)
- H (observation matrix) - indentity plus 1e-8 precision noise
- R (sqrt observation precision) - seems OK; projects correctly
- P0 (initial uncertainty) - prior uncertainty (OK)
- x (initial state) - prior mean (zero; OK)

Triangulation looks good, but posterior reconstruction is terrible.

wrong coordinate frame?

Spent hours troubleshooting math and tweaking parameters to understand the problem. Softening the prior seems to give better results. This doesn't tell us much, except that the data is doing its job correctly (which we already knew by observing the triangulation). Struggling to find the right combination of prior parameters.

Is the cubic prior simply not the right model?

Posted by Kyle Simek