answered
2015-10-07 20:15:04 +0200
Hi everyone (first post around here,yay!)
Another lousy runner here, multiple really bad injuries because I did some classic mistakes when starting running, and also still shitty programmer. This kind of app would save my feet, my back and my legs.
The accelerometer in the Jolla is quite sensible (but noisy at high frequencies). As juiceme mentioned more than one year ago, the first problem, which might be solved with some DIY, is to strap the phone on your body. Best places would be close to the center of gravity (so usually pelvis position/in the back - which is the position in the literature for posture analysis) or obviously on a leg. I might be tempted to say that we can use the I2C connectivity of the Jolla to have another sensor (there are some nice really small I2C 9 axis sensors from company like Invensense)... Or simply use a bluetooth accelerometer... but let's focus with only one sensor, in particular the accelerometer from the Jolla.
For me there are (at least) two way to see the analysis :
We can try to assess the quality of the running, in particular the "stability" of the running dynamic. There is a couple of things existing to analyze the walking stability and reflecting the physiological behaviour of the entire body (neuromuscular systems, joints, etc.). Those methods are however, from experience, very CPU and real-time-unfriendly (non-linear time series analysis). But we can look into the walking/gait analysis science to have some idea anyway :) !
We can try to get some indicators directly from the data by "simple" (I hate this word sometimes) kinematic analysis. I found some interesting resources, including one using a belt and one using a accelerometer on the leg.
First, https://peerj.com/preprints/370v1.pdf is more or less what we would like (right ?). The good news is page 21, where there is a difference on the patterns (from force plates, not accelerometers, however). From the same lab, in a poster (https://peerj.com/preprints/976/) they used those pattern differences to compute a time series from the data. It's quite easy to understand and the curves seems promising... However they can't reproduce the effect outside, this need some more investigations. I note they have a wonderful smooth graph of acceleration, don't expect this from the Jolla's sensor at full time-resolution... Some filtering will be needed (but it can be approached).
I also found a thesis (http://dash.harvard.edu/handle/1/14398557) which use a completely different approach, based on angles. I didn't read a lot, but as you can look from the page 37 to the conclusion, it might also be relevant.
There are also a lot of "basic indicators" that we can easily compute as the step cadency, etc.
Finally, the results will be very related to the subject/user, and some machine learning with a lot of subjects/users, with (or only) maybe an additional pre-training (for example "run on your heels for 10s", "run with exaggerated step length for 10s") might be necessary to have the app telling you directly "stop destroying your feet by running with sandals !" :-)
Actually I'm really busy those days, but if I have a bit of time, it might be fun to look into this project. I'll try to look more in details later and maybe record my next running session ! And I have definitively to check the videos to know more about this.
For feedback, I agree with the others, audio-feedback seems a great choice.
Sorry for the longpost and/or if I did mistakes and/or if I'm wrong. I'm sleepy and will probably come back on this message later ;)
