General Construction of Time-Domain Filter for Orientation Data

The basic idea behind this paper is that we can apply the filters we know
from DSP to orientation data. Clearly, there are no problems applying
filters on motion including only "Translation". However, rotations require
serious thinking, because the end results depend very much on the
representation we choose for the rotation. The authors show that
representing rotations as quaternions allows us to apply filters we know
with good results. The significance of that, is that now we can apply
kinds of filter we couldn't use before (for example, noise reduction,
Gaussian smoothing, or even edge detection to detect sharp changes in the
rotation)