Timeline for Creating features from raw accelerometer data
Current License: CC BY-SA 4.0
3 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jan 18, 2021 at 16:44 | comment | added | Matthew | Got it. Attempts to use LDA on data like this have used percentile-based features in my experience, so if you're not doing that it might be worth a try. Fourier features might be worth a try anyway, even though you're right about the nature of the activities - it could be that the act of balancing while standing (or getting tired after standing for a little while) induces some small movements at a higher frequency in the standing samples over the sitting. You might be right that the data isn't reliably separable (some of mine has been in the past, as I said) but you also might be surprised. | |
| Jan 18, 2021 at 16:12 | comment | added | Ben-Jamin-Griff | Mathew, thanks for your answer. I've looked over the dataset you suggested before but never saw the paper, this is very useful. I'm currently reducing the dimensions of my data down to a vector magnitude and calculating statistics from that. Each 'bout' is 30 seconds data sampled at 20hz and has a label for the activity. So far LDA seems like a good option. The issue with fourier features are that both sitting and standing arnt cyclic movements so I'm not sure the frequency content will help distinguish them. I think the issue is the data isnt reliably separable... | |
| Jan 18, 2021 at 14:22 | history | answered | Matthew | CC BY-SA 4.0 |