Abstract

The number of different methods available for filtering is vast. Which one of all available methods to use depends on several factors: the problem at hand, the type and quality of the estimates needed, the resources available, etc. In this talk different aspects of this design choice is discussed, based on the result in my recent Ph.D. thesis “Performance and Implementation Aspects of Nonlinear Filtering”. The results range from theoretical performance analysis and method comparisons to how to implement a parallel particle filter on a graphics card.

The main focus of the talk be on filter performance in terms of the root mean square error (RMSE) of the obtained the estimate. The RMSE is bounded from below by the Cramér-Rao lower bound (CRLB), which depends on both the model structure and the noise assumptions made. It will be shown how the CRLB depends on the intrinsic accuracy (IA) of the involved noise, where IA is a property of noise distributions quantifying information content. One of the results is that Gaussian noise constitutes a worst case assumption using this performance measure.