Particle filtering for systems with unknown noise probability distributions.

摘要

In recent years sequential Monte Carlo methods, or particle filtering, has attracted significant attention in the signal processing community. This has been primarily due to the flexibility and accuracy of particle filtering in resolving nonlinear non-Gaussian dynamic problems, where closed form analytical expressions are in general unavailable. Particle filtering approximates the a posterior density functions by discrete random measures, which are composed of particles and weights associated to the particles.; In this dissertation, we present a new class of particle filtering methods, called cost-reference particle filters (CRPFs), as opposite to conventional statistical-reference particle filters (SRPFs). The main feature of the new filters is that they are not based on any particular probabilistic assumptions while modeling the dynamics of the states and the observations as functions of the states. The underlying principle of exploring the spaces of the unknown states and parameters of the investigated system by particles remains the same as in standard particle filters. The difference rests on the generation of new particles and the update of their weights. Specifically, in the new family of sequential particle filters the statistical reference is substituted by a user-defined cost function that measures the quality of the state signal estimates according to the available observations. Another difference is the possibility of performing resampling in a decentralized way, or replacing the resampling by sorting, which makes these particle filters extremely attractive for hardware implementations.; Besides the theoretical development of specific methods in the new class, we provide experimental results that demonstrate the performance of the algorithms in the problem of maneuvering target tracking, and compare it with standard particle filters.; In the last part of the dissertation, some variations of the CRPF are discussed. The Unstructured CRPF estimates the unknown states as well as the transition functions. CRPF for system with conditional linearity is also investigated.