A collaborative method for dense motion estimation of fluid flows (back)

Description

Optical-flow methods dedicated to fluid motion have been recently introduced. These approaches have the main advantage to produce dense motion fields. However, as the underlying optimization are based on gradient descent techniques, the estimates are very sensitive to the initialization procedure. In particular, large displacement can only be recovered using a multiresolution strategy with the possible drawback of smoothing out small structure motions.
Fluid motion estimation techniques based on cross-correlation produce sparse motion fields. The estimation is in addition local and thus prone to erroneous spatial variability. However they are generally locally very robust to noise and may recover large displacement of some traceable features, that is to say features sufficiently contrasted and persisting over time on consecutive images.

To overcome the limitation induced by both optical-flow and correlation based methods, we propose to introduce an initialization scheme for a fluid dedicated dense motion estimation where the multiresolution scheme has been removed. The initial velocity field is estimated by minimization of a energy function composed of a 'div-curl' regularization and a data term. The latter constrains the dense estimated field to fit the sparse correlation vectors convoluted by bidimensional gaussian functions. The resulting vector field is then used to initialize the finest resolution of the fluid dedicated dense motion estimator introduced by [Corpetti et al., 2002].