Adjust kalman filter used for norfair tracker #9538
Conversation
✅ Deploy Preview for frigate-docs canceled.
|
|
Can you add a comment or some notes about this param and how it impacts things so we dont forget how we arrived at this value? |
|
Nick and I were discussing this a bit yesterday. A Kalman filter is like a car on the road with cruise control. Several parameters are used to configure it: a position variance, R and Q. The initial position variance is how confident we are in the car's position. It's inverse, so higher is less confident, and so the filter will rely on actually checking the position to get its bearings. Process Q is like hills and bumps on the road. A higher Q means it's more responsive to ups and downs. Then R is like a noise factor, how clearly it interprets the position from the readings vs how much it relies on predictions. Increasing Q would cause the Kalman filter to rely more on predictions rather than the actual measurements, reducing R would cause it to trust the measurements more. We reduced R from Norfair's default (4.0) down to 3.4. https://tryolabs.github.io/norfair/2.2/reference/filter/#norfair.filter.OptimizedKalmanFilterFactory |
* Adjust kalman filter * Add comment * Formatting
After testing I believe that the default values for the kalman filter are not optimal for Frigate.
This is with the current values:
we can see that the estimate (green) is consistently behind the actual object
with this PR:
we can see that the estimates are much more consistent with the actual detection.
I spent additional time testing this on other objects and it has performed well in tests