An application of sequential Monte Carlo samplers: an alternative to particle filters for non-linear non-gaussian sequential inference with zero process noise

Abstract

Particle filters are not applicable in sequential parameter estimation scenarios, ie scenarios involving zero process noise. Sequential Monte Carlo (SMC) samplers provide an alternative sequential Monte-carlo approximation to particle filters that can address this issue. This paper aims to provide a description of SMC samplers that is accessible to an engineering audience and illustrate the utility of SMC samplers through their application to a specific problem. The problem involves processing a stream of bearings-only measurements to perform localisation of a stationary tar get. The SMC sampler solution is shown to outperform an Extended and Unscented Kalman filter in nonlinear scenarios (as defined by a novel metric for nonlinearity that this paper describes). The SMC sampler offers a computational cost that is near-constant over time on average. Future work aims to investigate the utility of Approximate Bayesian Computation and apply the technique within a Simultaneous Localisation and Mapping context. (8 pages)