How to filter accelerometer data from noise

Question

Hello, attached is a plot of accelerometer data with 3 axis. The sudden bumps in the plot is the noise. I would like to get rid of them. So what filter should be used in this case ? If it is possible provide some pseudo code for it and explanation.

Regards, Levon

Answer 1

Looking at the data, you do not want any peeks that change the value by a certain amount (around 200, let us call this max_y_delta) in a certain time (5-15 perhaps, max_x_delta).

So as I'm not sure which structure your data has, I'll just assume it's 3 arrays data_array of floating point values that have a data point at every integer position. The solution I present is meant to be as simple as possible and you should try different values of max_x/y_delta to get good results. Even with the right values I'm sure there are much better solutions, but perhaps this one is good enough for you as a start.

max_x_delta = 10
max_y_delta = 200
for each of the 3 arrays
  for x = -1000...1000
    points_above_delta = 0
    average_value = 0
    for deltax = -max_x_delta/2...max_x_delta/2
      average_value += data_array[deltax]
      if abs(data_array[deltax] - data_array[x]) > max_y_delta
        points_above_delta++
      endif
    end for deltax
    average_value = average_value / max_x_delta
    if points_above_delta > max_x_delta/4
      for deltax = -max_x_delta/2...max_x_delta/2
        data_array[deltax] = average_value
      end for deltax
    end if  
  end for x

Note that this code has two downsides you might not want:

• The detection is very simple, there are also some peeks in your data that are meant to be there, so you might lose some of those.
• After detecting a peak, every value in the max_x_delta region around the peak is set to the average value in that region which will give you a straight line.

Answer 2

Try median filter: http://en.wikipedia.org/wiki/Median_filter, it can remove peaks, but not edges

Answer 3

It looks like you just want a low pass filter.

Answer 4

For measurement data where you can make an approximate model of what's physically going on, or what's physically likely, I'd suggest a Kalman filter. It's somewhat more complex than the other methods, but potentially gives cleaner output and/or better responsiveness.

Answer 5

Here is how I've done it http://levonp.blogspot.com/2010/10/how-to-filter-accelerometer-data-from.html