Live classifying human movement through machine learning | Hacker News Reader

Live classifying human movement through machine learning

blog.telenor.io

43 pointsjanjongboom5 comments

You have literally 6 features; you do not need to do dimensionality reduction here. Further, PCA is used generally before performing a linear regression model, which assumes independence among variables. SVM's are nonlinear and so don't care (though you specify both a linear kernel and the RBF kernel in your code?). If the data was very high dimension you could do PCA and then SVM in order to increase speed at the expense of accuracy - there is no free lunch. I think that's what you see in your results.

As a side note, SVM's and Random Forests are generally considered the best 'out of the box' models. So you chose a good one to experiment with!

Also, since you already wrote the analysis code in Python, why not write the front-end in Python (rather than Node) as well?

> Also, since you already wrote the analysis code in Python, why not write the front-end in Python (rather than Node) as well?

I had the node code laying around from a conference presentation I did, so it was the quickest way to gather some data while the others were looking into the machine learning part.

Then we just made the processes talk to eachother through filesystem, as you do at a hackathon..

> You have literally 6 features; you do not need to do dimensionality reduction here.

We do have six features, but they are not rotation invariant. I was hoping we could find the "axis of maximal motion" (or something along those lines) within each sample period, project the movement along that axis, and then do FFT to find what kind of bouncing around the test subject were doing. That didn't work out too well in practice, but it still sounds like the smart thing to do :-) What do you think?

I love posts like this, but I suspect another set of algorithms is going to win this race. Probably some combination of convnets and recurrent neural nets.

http://www.unnvision.com/pub/DDNN_ECCVW2014.pdf http://jmlr.org/proceedings/papers/v32/pinheiro14.pdf

https://www.moves-app.com/ also uses machine learning to classify walking, running, cycling etc. Not sure what algos but there are some pointers on the founder's website - http://www.cs.cmu.edu/~akyrola/

Reply on news.ycombinator.com