Learning reduced models for motion estimation on long temporal image sequences

This paper describes a sliding windows assimilation method, that allows estimating motion on long temporal image sequences, thanks to data assimilation techniques. The method splits the initial temporal window in sub-windows, on which reduced models are computed that allow to process images in quasi-real time. The method is quantified with twin experiments to demonstrate its potential for processing long-term satellite data. The main perspective is to replace the bases Ψ_ξ of the reduced models, which are obtained with a Principal Order Decomposition, by a fixed basis. In that case, even the first sub-window could be processed by a reduced model, in order to further reduce the computational requirements. Moreover, this fixed basis should be defined as satisfying optimality criteria, which translate properties on motion fields and image data. In that way, the method will be able to process long satellite sequences acquired over a full basin, as the Black Sea.