Accelerometer activity recognition

[Jakeun]

Hi everyone,

I am currently trying to recognize different activities such as walking, sitting and falling, using a watch which features a 3 axis accelerometer. Currently I have performed 1 minute tests of each activity 20 times but I am struggling on what to do with the data. Does anyone have any suggestions?

I have tried to FFT the data but do not have any decent peaks to establish a difference among the activities. Maybe I should apply a high or low pass filter?

ANY suggestions or advice would be greatly appreciated as I am really struggling with this.

Thank you

 

[AlephZero]

The first thing to do is plot graphs of the data from your tests (one test per graph) in the time domain, and spend some time looking at them. If I was doing something like this as a project at work, I would probably print out all the graphs and pin them up on the wall, and just leave them there for a few days to let my subconscious get to work on the data.

If you can't see any features that distinguish the situations, you probably need to think again about the concept. If you CAN see some distinguishing features, then you know what your mathematical signal processing needs to do. After that, if you attach a few of the plots here, somebody might have some ideas on how to do the math.

Presumably you don't know what was the orientation of the watch, so you might want to look at just the RMS acceleration (x^2 + y^2 + z^2)^(1/2) rather than the individual x y z components. Or you might want to see if the three components are closely correlated with each other, or independent. But without seeing the data, those are just random thoughts.