oduerr / bayes_cal_paper

bayes_cal_paper

Running on own data (in python)

To see how to use the calibration in R you can have e.g. a look at R/to_docs/calibration_3D_Glove.qmd all results discussed the paper are produced in R. The following example is taken from python/simple_calibration.ipynb

Data Collection

When recording, make sure to take data at many different rotations (you don't have to know the exact angle). Try to keep the device still during the recordings. You'll be recording the acceleration and it should look something like this.

accel = pd.read_csv('../data/IMU2.txt')
accel.head(5)

	accel_x	accel_y	accel_z
0	0.018066	-0.296875	1.040894
1	0.016357	-0.297974	1.037354
2	0.015137	-0.297607	1.032715
3	0.018188	-0.297485	1.034912
4	0.015015	-0.297607	1.039062

Calibration

You can run the calibration (ODR approximation) using

from cmdstanpy import CmdStanModel
dat = {'X' : accel, 'N' : accel.shape[0], 'D' : 3}
inits = {
    's' : [1,1,1],
    'b' : [0,0,0],
    'sigma' : 0.01
}
model = CmdStanModel(stan_file='../stan_models/calibration_ODR.stan')
fit = model.sample(data=dat, iter_warmup=10000, inits=inits)

You can then calibrate your readings using

b_mean = np.array(fit.summary().Mean[1:4])
s_inv = np.array(fit.summary().Mean[5:8])
accel_uncal = np.array(accel)
accel_cal = (accel_uncal - np.array(b_mean)) * s_inv

The effect of calibration is visible by looking at

$$ ||g|| = \sqrt{a_x^2+a_y^2+a_z^2} $$

before and after calibration. Since the device is not moving it should be close to 1 in all positions, they noise come from our far from perfect recording.

Reproducting the figures in the paper

Figure 1 (Priors): R/to_docs/prior.qmd

Figure 2 (2D Simulation) R/to_docs/calibration_2D_FullvsODR.qmd

Figure 3 (3D Simulation) R/larger_simulation/3D_Simu.R

Figure 5 (Calibration Sphere) R/larger_simulation/calibration_3D_Glove.qmd

Figure 6 (Effect of Calibration) R/larger_simulation/calibration_3D_Glove.qmd