Improved Mixture Representation in Real-Time Particle Filters for Robot Localization

Monte Carlo methods have been successfully adopted for robot localization thanks to their flexibility in distribution representation. However, these techniques are computationally expensive and can hardly perform at the incoming sensor data rate, when computation resources are limited. The Real-Time Particle Filter (RTPF) is an algorithmic solution conceived to make execution of a particle filter iteration feasible within time constraints by means of a mixture representation for the set of samples. RTPF requires an optimal balance of the contribution of each set to the mixture, whose computation, unfortunately, is quite difficult. In this paper, we provide a formal discussion of mixture representation by considering the weight mixture. We illustrate a novel solution for computing the mixture parameters based on the notion of effective sample size. This solution is less prone to numerical instability. Finally, we compare the proposed approach with the original RTPF algorithm through simulation tests and experiments.