Using models of dynamics for large displacement estimation on noisy acquisitions

Abstract : The paper discusses the issue of motion estimation on noisy images displaying large displacements, due to high velocity values. ``Noisy'' means that the data contain either missing acquisitions on isolated points, regions, frames or noisy measures. Assuming the dynamics is partially accessible from heuristics and modeled, the objective is to include this knowledge in the computation of the solution even if large displacements occur from one frame to the next one and if the data are noisy. This is performed by Data Assimilation techniques which simultaneously solve an evolution equation and an observation equation. The evolution equation includes the partial knowledge on the dynamics. The observation equation describes the transport of image brightness and is written in a non-linear form in order to better characterize large displacements. The assimilation method is a weak 4D-Var algorithm, in which each component of the Data Assimilation system is associated to an error. We prove that the observation covariance matrix can be used to discard the noisy data during the computation of the solution letting the evolution equation estimate motion from adjacent frames on these pixels. The method is quantified on synthetic data and illustrated on oceanographic satellite images.
Document type :
Reports
Complete list of metadatas

Cited literature [15 references]  Display  Hide  Download

https://hal.inria.fr/inria-00523625
Contributor : Dominique Béréziat <>
Submitted on : Wednesday, October 6, 2010 - 9:34:15 AM
Last modification on : Thursday, October 17, 2019 - 9:40:04 PM
Long-term archiving on : Friday, December 2, 2016 - 9:09:43 AM

File

RR-7408.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : inria-00523625, version 1

Citation

Dominique Béréziat, Isabelle Herlin. Using models of dynamics for large displacement estimation on noisy acquisitions. [Research Report] RR-7408, INRIA. 2010, pp.18. ⟨inria-00523625⟩

Share

Metrics

Record views

389

Files downloads

140