0
$\begingroup$

I'm a programmer/engineer that is getting into developing prototype wearable technology. I would like to be in a position to interpret raw accelerometer data and categorised the data into activities like running/walking/rest, etc.

The building and coding of the hardware and software is relatively straight forward, but I do not know the terminology and mathematical processes required in order to take the raw data and turn it into a training set that can be used for a real time system.

My question is this: What is a good DSP starting point to take a time based linear series and extract features so that they may be identified at a later stage?

At present, I have an tri-axis accelerometer device which samples at 50Hz with a range of -/+4g. I would like the capability of being able to readily identify activities based upon a training set, or machine learning process, if that's even possible?

Apologies that this isn't necessarily a specific DSP question, thanks for you assistance.

$\endgroup$
1
$\begingroup$

This seems very much like a fit for Hidden Markov Models (HMM). They are pretty powerful in handling sequence based input such as time series.

$\endgroup$
0
$\begingroup$

For periodic activities, such as running and walking, calculating the autocorrelation of sequences at least the length of several likely periods and reducing that to some number of maxima lags and local maxima values might provide a more useful feature vector than the raw data.

$\endgroup$
0
$\begingroup$

if you have a collection of feature templates, i s'pose cross-correlating the templates against the input sequence might show some of these features in the input. this is essentially what a "matched filter" does.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.