We have noticed markable improvements in move detection in accelerometer signal when we tested autocorrelation model. It’s similar to autoregressive (AR) model, but in this article we call this model autocorrelation.
The autocorrelation is defined as:
Where N is the total window size, N win is the AR window size, i is the AR shift, and Ncor is the number of steps.
In tests we have noticed that autocorrelation works a lot better than other methods, including FFT., though it requires significant calibration and tuning. See the graphs with different move states below.
The final plot depicted below is the result of running movement detection using autocorrelation on live production data collected from one of our customers baby,
The three colored axis show row accelerometer signal and the dotted line shows detected movement in moves per minute. The bottom graph shows state enumeration with different movement classification states detected. A thing to note that no-move detection is corresponds to flag=-2, which is an alarming condition to alert parents. In the sample below, the alarm have not been detected.
