13
Aug. 24, 2011
Smoothing Sensor Data with a Low-Pass Filter
If you search StackOverflow, there are a number of posts about Android sensor data being very jittery, and suggestions on implementing a smoothing algorithm. Well not many people have actually posted a smoothing algorithm, and those that I have seen tend to be fairly complex.
After some searching, I came to the Wikipedia entry for Low Pass Filter. There's a pseudocode algorithm that I adapted for Java which works splendidly. Not only that, but it's extremely simple. The algorithm essentially requires tracking only two numbers - the prior number and the new number. There's a constant, alpha, which affects the 'weight' or 'momentum' -- basically how drastically does the new value affect the current smoothed value. Here's the full implementation:
In my particular case, I used this to normalize raw accelerometer and magetometer sensor readings before calculating a compass bearing. Note that the input and output array elements are not sequential values, but completely separate dimensions (x,y,z) so e.g. each new x value is normalized against the smoothed x value, the new y with the smoothed y, etc.
This smoothing also has the curious effect of actually accelerating and decelerating the resulting visualization, e.g. the compass needle appears to have inertia and momentum if you quickly rotate the device. Just like a physical compass needle would. Neat!
Update: I've posted a follow-up, which explains how to take the low-pass output and get a compass rotation.
After some searching, I came to the Wikipedia entry for Low Pass Filter. There's a pseudocode algorithm that I adapted for Java which works splendidly. Not only that, but it's extremely simple. The algorithm essentially requires tracking only two numbers - the prior number and the new number. There's a constant, alpha, which affects the 'weight' or 'momentum' -- basically how drastically does the new value affect the current smoothed value. Here's the full implementation:
/*
* time smoothing constant for low-pass filter
* 0 ≤ alpha ≤ 1 ; a smaller value basically means more smoothing
* See: http://en.wikipedia.org/wiki/Low-pass_filter#Discrete-time_realization
*/
static final float ALPHA = 0.15f;
/**
* @see http://en.wikipedia.org/wiki/Low-pass_filter#Algorithmic_implementation
* @see http://developer.android.com/reference/android/hardware/SensorEvent.html#values
*/
protected float[] lowPass( float[] input, float[] output ) {
if ( output == null ) return input;
for ( int i=0; i<input.length; i++ ) {
output[i] = output[i] + ALPHA * (input[i] - output[i]);
}
return output;
}In my particular case, I used this to normalize raw accelerometer and magetometer sensor readings before calculating a compass bearing. Note that the input and output array elements are not sequential values, but completely separate dimensions (x,y,z) so e.g. each new x value is normalized against the smoothed x value, the new y with the smoothed y, etc.
This smoothing also has the curious effect of actually accelerating and decelerating the resulting visualization, e.g. the compass needle appears to have inertia and momentum if you quickly rotate the device. Just like a physical compass needle would. Neat!
Update: I've posted a follow-up, which explains how to take the low-pass output and get a compass rotation.
(Comments are closed)
13 Comments
This works very well, thanks! I'm implementing my own compass using the accelerometer and mag sensors. However, all examples I've found are incomplete in some way. Would you post your sensor handling code? It would be appreciated.
Dave - finally got around to posting an example which should clarify how the lowPass function is used. See my follow-up post.
Re: Smoothing Sensor Data with a Low-Pass Filter March 20, 2012 Paweł
output[i] = output[i-1] + ALPHA * (input[i] - output[i]);
?
Andrew - So the short answer is, no. See my follow-up post, which should clarify how the function is used.
You mean you're getting small numbers from the output of SensorManager.getRotationMatrix(), I presume? I don't remember but I think that value might be in radians.
Does anyone help me apply high-pass filter as well ?
As far as know, in order to remove noise I need to first apply "low pass filter " then "high pass filter" over it ( according to this post http://seattlesensor.wordpress.com/2013/01/01/accelerometer-sensor-data-processing/comment-page-1/#comment-98 , this way is the best way to remove the noise ) . Actually it would be great to learn the steps to remove the noise and have a proper set of acceleration data.
Regards,
I don't think you need to do a second filter if you use the technique I've detailed here. This smooths out all of the "noise" - if you were to graph the output it would be a pretty smooth curve. Increasing the value of alpha can make it smoother but it slows the rate of change.