Temporal smoothing particle filter for vision based autonomous mobile robot localization

Particle filters based on the Sampling Importance Resampling (SIR) algorithm have been extensively and successfully used in the field of mobile robot localization, especially in the recent extensions (Mixture Monte Carlo) which sample a percentage of particles directly from the sensor model. However, in the context of vision based localization for mobile robots, the Markov assumption on which these methods rely is frequently violated, due to "ghost percepts" and undetected collisions, and this can be troublesome especially when working with small particle sets, due to limited computational resources and real-time constraints. In this paper we present an extension of Monte Carlo localization which relaxes the Markov assumption by tracking and smoothing the changes of the particles' importance weights over time, and limits the speed at which the samples are redistributed after a single resampling step. We present the results of experiments conducted on vision based localization in an indoor environment for a legged-robot, in comparison with state of the art approaches.