Evaluation of a dispersion-based adaptive strategy using KinectTM and dynamic particle filter

Particle filters have been successfully applied to a variety of state estimation problems in recent years. In this paper we propose a novel and simple adaptive strategy to dynamically adjust sample sets in order to increase efficiency and drastically reduce computational time. The purpose of the dispersion-based adaptive particle filter (DAPF) is to quantify the level of particle distribution within the state space by determining the scattering distance. With this approach, the algorithm rapidly reduces the number of particles during the searching state when the dispersion decreases and quickly increases the number of particles during the monitoring state when the dispersion grows. Extensive experiments applied - but not limited - to RGB color tracking and mobile robot localization problems using KinectTM show that the DAPF approach significantly improves the computational performance over a generic PF with fixed sample set sizes and the adaptive technique named KLD.