Adaptive KLD sampling based Monte Carlo localization

The complexity of Monte Carlo Localization (MCL) mainly depends on the number of samples used for estimation and the KLD sampling method is a statistical approach to increasing the efficiency of MCL. The basic KLD sampling used fixed bin size whatever the distribution of samples is. This paper presents an improved KLD sampling method which adapts the bin size with the distribution. The idea of adaptive KLD sampling is to balance the approximation error and runtime with the dynamic bin size which is obtained by the space division though KD tree. The adaptive KLD sampling chooses big bin size when the state uncertain is high as well as small bin size when the samples focus on a little part of space. Simulation and experiment results show that adaptive KLD sample method yields improvement over the basic KLD sampling method.