You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hey,
I'm using your package right now to generate features for a dataset.
I have looked at the paper by Moe Nilssen et al. and tried to follow the steps for calculating step and stride regularity. However, I wonder why you still do the following calculation at the end:
@vanessabin it is mostly to help with comparison between devices and signal types.
If you calculate cross-correlation using xcorr(x, y, scale="coeff") then division by ac_lag0 wouldn't do anything as the data is already normalized. However, the other scale methods are not normalized and can't be directly compared.
Imagine one device that measures acceleration and another device that measures linear acceleration. The ac_d1 values for each device, for example, would be measuring different things and on different metrics (as one signal includes gravity, while the other doesn't). These values can't be directly compared against each other without first dividing each by their own "baseline" values to put them on a common metric
If you're looking to replicate the exact regularity values as calculated by Moe-Nilssen, then you could simply multiply the regularity values by the zero-lag autocorrelation peak value (ac_lag0) to reverse the process
Hey,
I'm using your package right now to generate features for a dataset.
I have looked at the paper by Moe Nilssen et al. and tried to follow the steps for calculating step and stride regularity. However, I wonder why you still do the following calculation at the end:
step_reg = ac_d1 / ac_lag0
stride_reg = ac_d2 / ac_lag0
Can you help me with this?
The text was updated successfully, but these errors were encountered: