Estimation de densités non paramétriques et multimodales par permutation de sous-particules. Application au suivi d'un ou de plusieurs objets synthétiques articulés

In this paper, we propose an original approach for sequential non-parametric density estimation defined in high-dimensional state spaces using the particle filtering framework. By exploiting conditional independences in the state space, we propose to swap independent subparticle sets to generate new sets that better sample this space. We integrate this approach into two versions of particle filter, i.e., partition sampling and annealed particle filter, to prove its efficiency. We compare it to classical approaches on synthetic articulated object density estimation problems, and show our approach reduces both estimation errors and computation times. MOTS-CLÉS : approximation de densité, filtrage particulaire, réseaux bayésiens.