Anisotropic fluid solver for robust optical flow smoothing

Abstract

Computational fluid dynamics provides the framework for explaining fluid motion. The methodology proposed in this paper applies partial differential equations such as Navier-Stokes for modelling the optical flow in image sequences displaying complex motion. The proposed robust stable fluid solver has the following components: robust diffusion, advection and mass conservation. We employ a robust diffusion kernel which combines the geometry preserving property of the heat kernel with an outlier rejection mechanism. The proposed methodology is applied on the artificially generated Von Karman flows, after considering additive noise, and onto the optical flow extracted from real image sequences.