(Tutoriel) Introduction générale aux méthodes particulaires en filtrage.
Résumé Cet exposé présentera une introduction générale aux méthodes particulaires pour le filtrage. Dans un premier temps, on rappellera le problème du filtrage et ses enjeux au travers de quelques applications dans des domaines aussi variés que la poursuite de cible, le traitement d'image et les mathématiques financières. On présentera ensuite les principales approches concurrentes ou complémentaires en spécifiant leurs forces et faiblesses. On exposera plus particulièrement le principe de base des méthodes particulaires et on précisera les points à améliorer de ces méthodes qui font actuellement l'objet de nombreux travaux de recherche. La présentation se terminera par un exposé des principaux résultats théoriques existants justifiant l'efficacité des méthode particulaires en filtrage.
Transparents
An introduction to particle methods in nonlinear filtering : fichier .pdf (560 ko)
Animation
illustration des étapes de prédiction / correction / sélection : fichier .gif animé (520 ko)
Particle filtering for positioning, with focus on positioning in wireless networks.
Abstract A general framework for positioning using particle filtering is described, including aspects as Rao-Blackwellization, integer implementations and robustness issues. A survey of recent applications is presented, covering underwater and surface vessels, airborne systems and vehicles. Particular attention is paid to positioning of mobile terminals in wireless networks, utilizing time-difference of arrival and other network measurements.
Transparents
Particle filtering for positioning, with focus on positioning in wireless networks : fichier .pdf (3048 ko)
(Tutoriel) Algorithmes pour le filtrage particulaire : un état de l'art.
Résumé Dans cet exposé tutoriel, je décrirai un ensemble de méthodes génériques permettant de construire des algorithmes particulaires efficaces :
Transparents
Sequential Monte Carlo methods. A review : fichier .pdf (144 ko)
Recalage de centrale inertielle par gravimétrie.
Résumé
Transparents
Gravimetry-aided inertial navigation : fichier .ppt (2384 ko)
Target tracking and guidance using particles.
Abstract Most developments in "modern guidance laws" have been based on linear-Gaussian (LG) assumptions and the application of linear-quadratic-Gaussian (LQG) theory. However, in real-world problems the LG assumption is often grossly violated and a quadratic cost function may be inappropriate. In particular, the certainty equivalence assumption may be invalid.
The stochastic control problem will be considered in the context of particle filters. A guidance law that makes good use of the sample set from a particle filter is proposed. The operation of the filter / guidance law will be demonstrated and analysed for a two object simulation problem.
Transparents
Target tracking and guidance using particles : fichier .pps (608 ko)
Particle filters for target tracking.
Abstract Some models will be presented that do not fit neatly into the framework associated with Kalman filtering, currently used extensively in target tracking. These models will be used to show that particle filtering can capitalise on model fidelity and so offer the potential to improve performance. Some illustrative examples will be used to demonstrate applications of particle filtering within this context.
Transparents
Particle filters for target tracking : fichier .ppt (1160 ko)
Demos
I can't (yet) put the demos on the web, but I do intend to be able to do that at some point.
(Survey) Pixels and particles : sequential Monte Carlo for image analysis.
Abstract Since the seminal work by Isard and Blake in 1996, particle filters have become a popular set of tools in the toolbox of the computer vision / image analysis practitioners. In particular, they are now routinely used to track all sorts of entities in videos. The first part of the presentation will offer a (partial) survey of this six-year history, along with
Transparents
Pixels and particles : sequential Monte Carlo for image analysis : fichier .pps (3936 ko), fichier .zip incluant les vidéos (14136 ko)
Using multi-modality to guide visual tracking.
Abstract Visual tracking is notoriously difficult. This is due to large number of factors, including changes in appearance due to changes in lighting and illumination, occlusions, clutter, high dimensional state spaces, and many more. The robustness of tracking can be improved by combining the information available in complementing modalities. This will be illustrated in two settings. In the first a contour based visual likelihood is combined with a sound source localisation likelihood. The configuration of the system is such that the sound likelihood gives information about the direction of the sound source, which can be translated to a horisontal position in the image plane. Thus if the person being tracked is speaking, ambiguities in the horisontal direction can be resolved by the sound likelihood. The resulting estimate is then further refined by the visual likelihood. The combination of sound and vision likelihoods can also be used to focus the attention on speakers as they alternate in conversation. In the second setting two visual modalities, namely colour and motion, are used for head tracking. During periods of motion colour information is unreliable due to changes in pose and illumination. Tracking then mainly relies on the motion likelihood. Conversely, when the object is stationary or near-stationary motion information disappears, and localisation relies on the colour information. It is also shown how the colour model can be adapted during periods of motion to better capture the characteristics of the individual.
Transparents
Using multi-modality to guide visual tracking : fichier .ppt (2120 ko)
Image sequence based particle filter for point tracking.
Abstract The approach we investigate for point tracking combines within a stochastic filtering framework a dynamic model relying on the optical flow constraint and measurements provided by a matching technique. Focusing on points whose motion may only be described by a local parametric model, the tracking system we propose is composed of a nonlinear dynamic equation and a Gaussian observation equation. In that context, the tracker is built from a conditional particle filter whose optimal importance function is known. Since we focus on on the case where the system depends on the images, such a dependency has to be taken into account through a conditioning with respect to the image sequence data. This conditional tracker authorizes to significally improve results in some general situation. In particular, such an approach allows us to deal in a simple way with the tracking of points following trajectories with abrupt changes.
Transparents
Image sequence based particle filter for point tracking : fichier .ppt (616 ko), fichier .zip incluant les vidéos (7408 ko),
Videos
Particle filtering for joint data-channel estimation in fast fading channels.
(travail commun avec Didier Le Ruyet, Gilles Rigal et Han Vu-Thien)
Abstract
fichier .pdf
Transparents
Particle filtering for joint data-channel estimation in fast fading channels : fichier .ppt (512 ko)
Quelques applications du filtrage particulaire pour les communications numériques.
(travail commun avec Elena Punskaya, Bill Fitzgerald et Xiaodong Wang)
Résumé De nombreux problèmes de communications numériques peuvent se mettrent sous la forme d'un problème de filtrage non-linéaire. Il est tentant dès lors de vouloir appliquer le filtrage particulaire à ces problèmes. Je présenterai plusieurs applications du filtrage particulaire dans ce contexte :
Transparents
A few applications of particle filtering to digital communications : fichier .pdf (376 ko)
Retour à la page de François Le Gland