Numerical Fitting-based Likelihood Calculation to Speed up the Particle Filter
For real-time particle filtering in rich-observation applications, this method reduces computational bottleneck, but is incremental as it builds on existing fitting techniques.
The paper proposes a numerical fitting approach to speed up particle filter likelihood calculation by constructing a Likelihood PDF from a small number of fulcrums, reducing computation and enabling real-time filtering. Simulations on 1D and multi-dimensional tasks (robot localization, visual tracking) show validity.
The likelihood calculation of a vast number of particles is the computational bottleneck for the particle filter in applications where the observation information is rich. For fast computing the likelihood of particles, a numerical fitting approach is proposed to construct the Likelihood Probability Density Function (Li-PDF) by using a comparably small number of so-called fulcrums. The likelihood of particles is thereby analytically inferred, explicitly or implicitly, based on the Li-PDF instead of directly computed by utilizing the observation, which can significantly reduce the computation and enables real time filtering. The proposed approach guarantees the estimation quality when an appropriate fitting function and properly distributed fulcrums are used. The details for construction of the fitting function and fulcrums are addressed respectively in detail. In particular, to deal with multivariate fitting, the nonparametric kernel density estimator is presented which is flexible and convenient for implicit Li-PDF implementation. Simulation comparison with a variety of existing approaches on a benchmark 1-dimensional model and multi-dimensional robot localization and visual tracking demonstrate the validity of our approach.