I have a semi-regularly (15fps to 60fps) sampled source of position and orientation to which I am using dead reckoning to fill in gaps.

However, I would like to smooth it further to eliminate any steps due to dead reckoning error.

The x,y,z can easily be smoothed with a small (3pts?, 5pts?) moving average filter. However, I am unsure of how to approach smoothing the orientation.

I have access to Axis angle representations and a decent Quaternion class that I could possibly build a moving average filter with. It has multiply (operator *) and power (operator ^) which in theory would allow me to effectively average a set of rotations.

Would filtering an axis angle representation work if I simply filtered the 'x','y','z' and 'a' channels and then normalized it?