Make gyro bias clip relative to any previously known bias. #23
Conversation
Prior to this commit, if the bias of a sensor exceeded the biasClip parameter, it would be unable to detect rests, even if you gave VQF a known estimate using VQF::setBiasEstimate or similar.
|
Hm. No, biasClip is used elsewhere in an absolute sense. This is either incomplete, or wrong. |
|
Sorry for the late reply. I somehow lost track of this. In the future, feel free to ping me after some time. :) I think, in general, the change you suggested would be a valid design decision (as would be to just remove this additional condition and only consider The rationale behind the absolute limit of 2 °/s for bias estimation was that
The reason to also include this threshold in the rest detection logic is to prevent some weird corner cases where the angular rate is surprisingly constant (i.e., some very smooth constant rotation like on a rotating table). I think, generally, the best strategy would be to work on the calibration and ensure that the gyroscope bias that reaches VQF varies by less than 1 °/s. If this is not possible, I think modifying VQF locally is a good option. Unfortunately, what I won't do is to modify the behavior of the VQF algorithm itself since this would invalidate the extensive validation done in the publication. For that reason, I will close the merge request for now, but I'm happy to discuss this further. |
Just ran into this: for some reason the native calibration tool for my sensor is doing a poor job of calibrating the gyro bias (or the sensor isn't taking the bias into account when producing quaternions?), so the default gyroClip value was causing the algorithm to never detect a rest. Setting the gyro bias via
VQF::setBiasEstimatedidn't help, which led me to the realization that the gyroClip doesn't take into account any known bias.It personally makes more sense to me to make gyroClip relative to any already known bias than it does to raise gyroClip in this situation. Maybe there's a reason not to do this, though.
(aside: i don't have a good environment to test the python or matlab implementations right now, so you should probably make sure they still interpret/compile before accepting this PR.)