Coupling Reduced Models for Optimal Motion Estimation

The paper discusses the issue of motion estimation by image assimilation in numerical models, based on Navier-Stokes equations. In such context, models' reduction is an attractive approach that is used to decrease cost in memory and computation time. A reduced model is obtained from a Galerkin projection on a subspace, defined by its orthogonal basis. Long temporal image sequences may then be processed by a sliding-window method. On the first sub-window, a fixed basis is considered to define the reduced model. On the next ones, a Principal Order Decomposition is applied, in order to define a basis that is simultaneously small-size and adapted to the studied image data. Results are given on synthetic data and quantified according to state-of-the-art methods. Application to satellite images demonstrates the potential of the approach.