Noise densities and random walks¶
White noise densities and bias random walks of accelerometer and gyroscope,
read off the Allan deviation as in Kalibr and allan_variance_ros. Background:
Kalibr IMU noise model,
El-Sheimy et al., 2008.
Recording¶
At least three hours on a solid surface in a quiet room. Fans, compressors and footsteps show up in the curve.
Compute¶
/imu/data: 4320000 samples over 10800.0 s (400.0 Hz)
accelerometer_noise_density: 2.027539e-3
accelerometer_random_walk: 9.985073e-4
gyroscope_noise_density: 3.053876e-4
gyroscope_random_walk: 2.169208e-4
update_rate: 400.00 Hz
The curves go to allan.csv, one row per cluster time
(tau,accel_x,accel_y,accel_z,gyro_x,gyro_y,gyro_z,pairs). On log-log axes
the white-noise region falls with slope −½ and the random-walk region rises
with +½. Bumps or an early rise indicate disturbance.
When it refuses¶
Recordings under three hours give a warning. Without a usable +½ region the fit stops:
error: the Allan deviation of the accelerometer x axis has no clear slope +0.5 region
(0 point(s) spanning 0.0x in tau, need 4 spanning 1.5x)
Record longer, or raise noise.cluster.max up to a tenth of the recording:
Datasheet values¶
Angle and velocity random walk give the white-noise densities after unit conversion. Bias instability is not the bias random walk. The Kalibr wiki describes an estimate. Datasheet values are optimistic.
Pelorus covariances¶
imu noise and imu show print the Pelorus imu.covariance_* block, and
--pelorus pelorus.rl writes it:
imu {
covariance_gyroscope = 0.003730463
covariance_gyroscope_bias = 0.001882185
covariance_acceleration = 0.1644366
covariance_acceleration_bias = 0.03988067
}
These are discrete-time variances, (density · inflation)² · rate. The Allan
deviation of a still sensor at constant temperature underestimates the noise
on a moving robot, so noise.inflation (10) scales the densities first. The
Kalibr wiki suggests 10× or more for MEMS parts. Check the factor on a
recording with a known trajectory.