A new algorithm for outlier rejection in particle filters

Filtering algorithms have found numerous application in various fields. One of the main factors that affect the performance of filtering algorithms is when the instrument recording the observations is faulty and yields observations which are outliers, that subsequently degrade the performance of the filter. A standard procedures to deal with this issue is to reject any measurement that is at least three standard deviations away from the predicted measurement. This method works very well for linear Gaussian estimation. For particle filter which does not require any Gaussian assumptions, the aforementioned noise rejection procedure yields poor performance. In this paper, we present a new outlier rejection procedure for particle filters that uses the theory from kernel density estimation and probability level sets. The proposed solution does not impose any constraint on the type of noise or the system transformation, and consequently the particle filter realizes its full potential. Simulation examples are presented in the end to show that our proposed algorithms works better than conventional outlier rejection algorithm.