email: Elise.Arnaud@irisa.fr address: IRISA / INRIA Rennes Campus Universitaire de Beaulieu 35042 Rennes Cedex - FRANCE ph: +33 2.99.84.71.67 fax: +33 2.99.84.71.71 secretary: +33 2.99.84.72.52 (Stéphanie Lemaile)
|
![]() |
In 2001, I graduated with a five year engineering degree from the applied mathematics department of INSA (National Institut of Applied Sciences) at Rouen, France. During the final year of this degree, I obtained a DEA specialized in image processing, at the university of Rouen. In October 2001 I started a PhD in the VISTA research project at IRISA Rennes. My supervisor is Etienne Mémin.
My research concerns the use of filtering methods for tracking in image sequences. In the case of single object tracking, the Kalman filter gives an analytic recursive solution if the model is linear and Gaussian. In the non-linear and/or non-Gaussian case, sequential Monte-Carlo methods (also called particle filters) provide an approximation of the expected a posteriori distribution in terms of a finite sum of Diracs, centered in elements of the state space named particles.
First, a new conditional formulation of classical filtering methods has been proposed which is dedicated to image sequence based tracking. These conditional filters allow solving systems whose measures and state equation are estimated from the image data. In particular, the model that is considered for point tracking combines a state equation relying on the optical flow constraint and measures provided by a matching technique. Based on this, two point trackers have been derived. The first one is a linear tracker well-suited to image sequences exhibiting global dominant motion. The second one is a nonlinear tracker, implemented from a conditional particle filter. This allows tracking of points whose motion may be only locally described.
Secondly, the interest of optimal importance function use has been studied. This function has a crucial role in the diffusion step of the particle filter algorithm. In that framework, a model dedicated to point tracking in cluttered background has been built. It allows taking into account several possible measures by considering a likelihood modeled as a Gaussian mixture. I has been demonstrated that the optimal importance function can be used with such a model. This system significatly increases result quality in case of measure ambiguity.
We currently work on the use of hybrid filters (adapted to conditionally Gaussian models) for deformable objet tracking.