Robust techniques for multiple target tracking and fully adaptive radar = Técnicas robustas para seguimiento de múltiples blancos y radar adaptativo

Abstract

This Ph.D. thesis is concerned with the development of robust methods for high-dimensional recursive Bayesian filtering and sensor management in dynamic systems. Many problems in science and engineering require the estimation of the state of a system, that changes over time, using a sequence of available noisy measurements. Recursive Bayesian filters sequentially compute at each time step the probability density function (PDF) of the state of a dynamic system given all the received sensor measurements. This posterior PDF includes all the information of interest about the state for estimation purposes. However, except for a very limited class of models, the posterior PDF cannot be generally computed in closed form, and approximations, such as particle filters, are necessary. Unfortunately, the performance of these approximations generally severely degrades when the dimension of the space in which the state of the system takes values is high, an effect which is commonly referred to as the curse of dimensionality. Part I of this thesis focuses on the development of different particle filtering techniques which can robustly tackle the filtering of high-dimensional states. A useful strategy to overcome the curse of dimensionality is to consider a partition of the state space, so that samples from each component of the partition can be drawn independently in a particle filter. Following this strategy, the auxiliary parallel partition (APP) method is proposed, which overcomes limitations of previous partitioned particle filters in the literature by considering the use of auxiliary particle filtering. A second filter, which additionally incorporates a component-resampling (target-resampling) stage, is also presented. This filter, the target-resampling APP (TRAPP) is shown to be useful when the dimension of the state space is high, at the cost of a loss in the diversity of the samples of each component. Thus, a third method is considered, the adaptive TRAPP (ATRAPP), which adaptively decides if target-resampling is needed in each component. This makes ATRAPP a robust algorithm, with a reliable performance regardless of the dimension of the state space, performing target resampling when necessary and favoring sample diversity when possible. An alternative strategy to beat the curse of dimensionality is to make use of multiple filtering, where the marginal posterior PDF of each component of the state is individually estimated using a different filter. Multiple filters, however, require of a marginalization procedure, which generally also needs to be computed in an approximated form. The inclusion of auxiliary particle filtering along with a first-order approximation to the marginalization procedure in the multiple auxiliary particle filter (MAPF) is detailed. In addition, the sigma-point multiple particle filter (SP-MPF) is presented, which, making use of sigma-point integration methods, computes a second-order approximation to the required marginalization procedure. Finally, auxiliary filtering is also considered within this setting in the sigma-point multiple auxiliary particle filter (SP-MAPF). The presented algorithms are shown to have an outstanding performance with respect to previous methods in the literature in a multiple target tracking scenario. It is often the case in which the received measurements depend on some parameters of the sensors which can be tuned. The posterior PDF of the state in such cases therefore depends on the selected sensor parameters, so that they need to be carefully chosen to maximize performance. This sensor-management problem is the focus of Part II of this thesis. Recent approaches to the sensor management problem, such as the fully adaptive radar (FAR), are inspired by the neuroscience approach to decision making and cognition. The FAR framework gathers in a simple and compact architecture the main concepts of this novel approach. Implementations in the literature of the FAR framework for the problems of single target tracking and simultaneous single target detection and tracking are first reviewed, and the specialization of the FAR for the problem of multiple target tracking with a fixed and known number of targets is presented. In the final part of this thesis, the reviewed FAR implementations in the literature for the problems of single target tracking and simultaneous single target detection and tracking are shown to suffer robustness issues when applied to difficult scenarios. These robustness issues are first thoroughly characterized and alternative robust novel methods within the FAR framework are presented. The proposed methods are shown to reliably perform in these difficult scenarios.