What is the largest public wearable accelerometer dataset?

Question

I'm looking for public datasets of people wearing a device with an accelerometer (and potentially other sensors eg gyroscope or magnetometer). What are some of the largest available datasets like that?

This has obvious applications to machine learning: a good dataset will help develop good models for activity recognition and event detection from sensor data, same as the effect ImageNet / COCO / YFCC100m datasets have had in the visual field. Sadly, I think the very large datasets in this field are all private.

For my purposes, I don't care what people are doing; a totally random sample of activity is okay (but a broader sample is better than a specific sample). I also don't care where the device is (wrist, pocket etc) or whether it is a smartphone or some other device (watch, actigraph, IMU etc). Finally, I don't care if the data is labelled / annotated.

I do however want the largest possible size: as many different people as possible, and as many total hours of recording as possible.

What I've found so far...

A few collections of datasets: https://arxiv.org/pdf/1707.03502.pdf and http://mobilize.stanford.edu/data-sources/

Some specific datasets:

• UK Biobank has recordings from 100k people x 24 hours each. I believe these are actigraph (1-minute resolution) not raw accelerometer data (can anyone confirm?). It is also not open.

• NHANES 2003 7k people x 7 days each. Definitely actigraph, no raw accelerometer data.

• LTMM 71 people x 72 hours each

• PAMAP2 9 people x ~1 hour each

• MHEALTH 10 people x ~15 minutes each

• UCI-Smartphone

Answer 1

You can check these additional ones:

HAR: http://archive.ics.uci.edu/ml/datasets/Human+Activity+Recognition+Using+Smartphones It is collected by attaching a smart-phone and it has accelerometer and gyroscope data. It has 30 participants conducting 6 different activities. The acceleration is at 50 Hz and they are using 3D accelerometers.

USCHAD: https://dl.acm.org/citation.cfm?doid=2370216.2370438 Have a look at this paper. Here you will find the data as well.

It has 14 subjects (7 male, 7 female) and there are 12 different activities. Additionally, it has accelerometers and gyroscopes, and it has a ground truth labeling by an observer standing nearby. They are sampled on 100Hz. Again they are using 3D accelerometers.

I would strongly suggest feature-engineering by defining time-windows where you segment the time-series data. After defining the small windows of a few seconds, you can have one activity per window. Then you can easily extract some time-dependent features such as mean, correlation, acceleration, orientation, etc.. Also, Fast Fourier transform on these small windows can be applied so you can extract many frequency domain features like energy, coherence, etc.