- refactor updateQuaternion() to compute the yaw jacobian directly (respecting the rotation sequence determination)
- fuseHeading()/fuseYaw321()/fuseYaw312() helpers are eliminated and now mag heading fusion and EV yaw fusion compute the innovation in place
- clear up logic for performing zero innovation heading fusion when quaternion variance exceeds threshold (no more _is_yaw_fusion_inhibited flag manipulation)
- when at rest continue fusing last static heading with very low variance even if other heading sources are active
As they are always moving horizontally, the check doesn't make sense
for fixed-wings.
Also don't run the check while on ground to prevent getting a failed
check during pre-takeoff manipulation.
When wind is already estimated, we don't reset the states using airspeed
data, so it could be that the fusion fails if the airspeed isn't
consistent with the filter (test ratio > 1). In this case, don't start
the fusion.
When wind isn't already estimated, the wind states are reset using
airspeed so the fusion can start regardless of the current test ratio.
split the fusion process into:
1. updateAirspeed: computes innov, innov_var, obs_var, ...
2. fuseAirspeed: uses data computed in 1. to generate K, H and fuse the
observation
- all sources of optical flow publish sensor_optical_flow
- sensor_optical_flow is aggregated by the sensors module, aligned with integrated gyro, and published as vehicle_optical_flow
Co-authored-by: alexklimaj <alex@arkelectron.com>
The noise spectral density, NSD, (square root of power spectral density) is a
continuous-time parameter that makes the tuning independent from the EKF
prediction rate.
NSD corresponds to the rate at which the state uncertainty increases
when no measurements are fused into the filter.
Given that the current prediction rate of EKF2 is 100Hz, the
same tuning is obtained by dividing the std_dev legacy parameter by 10:
nsd = sqrt(std_dev^2 / 100Hz)
The baro observation noise parameter is often over-estimated in order as
a measure to mitigate temporary offsets in the readings due to wind
gusts or poor pressure compensation tuning. The side effect is that the
innovation sequence monitoring based on normalized innovation struggles
to detect an offset in the state because the innovation isn't
statistically significant.
To counter this issue, a simpler check is added to trigger the process
noise boost when the innovation has the same sign for a long period of
time.
Calculating K(HP) takes less operations than (KH)P because K and H are
vectors.
Without considering the sparsity optimization:
- KH (24*24 operations) is then a 24x24 matrix an it takes
24^3 operations to multiply it with P. Total: 14400 op
- HP (24*(24+24-1) operations) is a row vector
and it takes 24 operations to left-multiply it by K. Total:1152 op
Daniel Agar