Beta 1


Title Filtering in hybrid dynamic Bayesian networks
Author Andersen, Morten Nonboe (Intelligent Signal Processing, Informatics and Mathematical Modelling, Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark)
Andersen, Rasmus Ørum
Supervisor Hansen, Lars Kai (Intelligent Signal Processing, Informatics and Mathematical Modelling, Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark)
Institution Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark
Thesis level Master's thesis
Year 2003
Abstract In this thesis we describe the use of the Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), generic Particle Filter (PF, a.k.a. condensation, survival of the fittest, bootstrap filter, SIR, sequential Monte Carlo, etc.), Particle Filter with MCMC steps (PFMC), Particle Filter with EKF proposal (PFEKF) and MCMC steps (PFEKFMC), Particle Filter with UKF proposal (PFUKF) and MCMC steps (PFUKFMC) in theory as well as in a practical framework. We present pseudo-code (from Merwe00) for all algorithms and implement the filters in a Dynamic Bayesian Network (DBN) framework using Matlab. Furthermore, we demonstrate and compare the implementations on a simple one-dimensional state estimation problem, a more complex simulation of a watertank system and finally on a real-life problem in which we use a cyberglove to infer the angle, angular velocity and angular acceleration of a single fingerjoint during movement and use these variables as hidden nodes in a 2T-DBN with EMG measurements from the lower arm as observations. Furthermore, we show how the filters differ theoretically as well as practically and when and how their strengths and weaknesses become visual. Finally, we conclude which filters are superior under different conditions and in different practical scenarios. Theory and implementation is based on the theory and pseudo-code presented in Merwe00. Dansk abstract: I dette eksamensprojekt præsenteres the Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), generic Particle Filter (PF, a.k.a. condensation, survival of the fittest, bootstrap filter, SIR, sequential Monte Carlo, etc.), Particle Filter med MCMC steps (PFMC), Partice Filter med EKF proposal (PFEKF) og MCMC steps (PFEKFMC), Particle Filter med UKF proposal (PFUKF) og MCMC steps (PFUKFMC). En teoretisk gennemgang af filtrene afsluttes med opskrivning af pseudo-koden (fra Merwe00) for en implementering af det enkelte filter. Desuden implementeres de forskellige filtre i et Dynamisk Bayesiansk netværks (DBN) framework vha. Matlab og vi demonstrerer og sammenligner teknikkerne vha. et simpelt en-dimensionalt state estimations problem og en større og mere kompleks simulation af et vandtankssystem. Endelig anvendes udvalgte filtre på et problem fra den virkelige verden, hvor vi anvender en cyberglove til at inferere vinkel, vinkelhastighed og vinkelacceleration for et enkelt fingerled i bevægelse og anvender disse variable som skjulte knuder i et 2T-DBN med EMG målinger fra underarmen som observationer.
Series IMM-Thesis-2003-32
Keywords Hybrid Bayesian Networks; Dynamic Bayesian Networks; Particle Filtering; Extended Kalman Filter; Unscented Kalman Filter; Markov Chain Monte Carlo
Fulltext
Original PDF imm2435.pdf (2.37 MB)
Admin Creation date: 2007-06-11    Update date: 2007-09-19    Source: dtu    ID: 200684    Original MXD