Computationally efficient algorithms for non-linear target tracking problems
Author(s)
Yaqoob, Muhammad Moeen
Type
Thesis
Abstract
Target tracking concerns the processing of a sequence of noisy sensor measurements to estimate the state of a target. These measurements are expressed as a noisy function of the state according to a measurement model while the evolution of the state with time is described by a system model. Among the recursive Bayesian estimators, the Kalman filter gives the exact solution to the tracking problem if both the underlying models are linear and Gaussian. However, a number of sub- optimal filters can be used for more general non-linear tracking problems, including the well known bearings-only and range-only tracking problems. The traditional sub-optimal moment-matching filters rely on linear-Gaussian approximation of the underlying non-linear model and are, therefore, not very accurate. On the other hand, particle filters, based on direct approximation of conditional densities of the state, are versatile and accurate but computationally much more expensive. We claim that the reason for poor performance of the moment-matching methodology is the manner in which it is carried out. The aim of this research is, therefore, to develop new moment-matching filters that can be used in situations where conventional trackers give poor estimates or make excessive computational demands. We introduce two moment-matching tracking algorithms, named the ‘shifted Rayleigh filter’ and the ‘quadrature range-only filter’. These filters provide efficient sub-optimal Bayesian solution to highly non-linear tracking problems, associated with bearings-only and range-only measurements. Unlike other moment-matching methods, both these filters are based on exact calculation of the mean and covariance of the non-normal updated density of the target state, given a normal approximation to the previous state. Simulations show that the algorithms either outperform or are competitive with standard moment-matching methods. In particular, the shifted Rayleigh filter achieves the accuracy comparable to that of particle filters, while reducing the computational overhead by several orders of magnitude.
Version
Open Access
Date Awarded
2008
Advisor
R B, Vinter
Sponsor
Data and Information Fusion Defence Technology Centre (DIF-DTC); Ministry of Defence and General Dynamics UK.
Publisher Department
Department of Electrical and Electronic Engineering
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)