Highly accurate approaches to motion estimation – quantifying fluid dynamical transport phenomena

applications to quantifying fluid dynamical transport phenomena. Extensions
to the motion model will be presented that make the accurate estimation of
motion and other transport parameters feasible.
In the presence of noise, it is important to limit the number of model
parameter. Otherwise overfitting will occur. A novel approach to model
selection performs statistical tests on the residual of a local estimator to
evade the problem of overfitting.
In a local framework, due to the spatio-temporal grey value structures, the
parameters of the motion models cannot be attained accurately in all
regions. Therefore, a statistical confidence measure will be presented, that
makes it feasible to discard inaccurate estimates.
In a post-processing step, the resulting holes in the estimates are filled
with an approach to inpainting or by regularizing with typical vector fields
learned from PCA.
The proposed techniques result in a statistically sound, highly accurate
approach to motion estimation that is straightforward to parallelize and
thus speed up.