An Efficient Adaptive Sampling Algorithm for Particle Filtering of the Hierarchically-Modeled Object
Takashi BANDO, Tomohiro SHIBATA, Mikio SHIMIZU, Shin ISHII
This paper presents a novel method for estimation of high-dimensional state variables by means of particle filtering (PF). One of the major drawbacks of PF is that a large number of particles is generally required for accurate estimation of state variables lying in a high-dimensional space, whose maintenance is time-consuming. In many applications, the high-dimensional state variables can be divided into two groups according to the hierarchy that an object model possesses; one group (higher layer) is easily integrated out analytically from the posterior densities, and we have only to carry out PF for the other group (lower layer) whose state space is reduced from the whole state space. In this paper, we propose a novel proposal distribution in the lower layer, which is a mixture of approximate prediction densities computed in the higher and lower layers. An adaptive determination of the mixture ratio, which implements a mutual interaction between the layers as a mixture state transition model in PF, is realized by means of on-line EM algorithm for adapting to complicated real environments. The effectiveness of the proposed method is demonstrated by computer simulations of head pose estimation of a car driver.