The particle filter can give the approximate solutions to the non-linear non-Gaussian model of mobile robot global localization.However,if the new measurements appear in the tail of the prior or if the likelihood is too peaked in comparison to the prior,the conventional particle filter can degenerate and make localization fail.We present a novel algorithm that combines an importance sampling with central difference filter (CDF).The posterior pose state density is represented by Gaussian mixture model (GMM) that is recovered from the weighted particle set of the measurement update step by means of a weighted expectation maximization (WEM) adaptive clustering algorithm,which based on the kd-trees.Experimental results show that this new approach has an improved localization accuracy and reduceds computational complexity.