Feature-based annealing particle filter for robust motion capture
Tutor/a - Director/a
Estudiant
López Méndez, Adolfo
Tipus de document
Projecte Final de Màster Oficial
Data
2009
rights
Accés obert
Editorial
Universitat Politècnica de Catalunya
UPCommons
Resum
This thesis presents a new annealing method for particle filtering aiming at body pose
estimation. Particle filters are Monte Carlo methods commonly employed in non-linear
and non-Gaussian Bayesian problems, such as the estimation of human dynamics. However, they are ine±cient in high-dimensional state spaces. Annealed particle filter copes
with such spaces by introducing a layered stochastic search. Our algorithm aims at generalizing and enhancing the classical annealed particle filter. Diferent image features
are exploited in a sequential importance sampling scheme to build better proposal distributions from likelihood. This technique, termed Feature-Based Annealing, is inferred
from the required function properties in the annealing process and the properties of the
weighting functions obtained with common image features in the field of body tracking.
Comparative results between the proposed strategy and common annealed particle filter
are shown to assess the robustness of the algorithm.
