The efficiency of stochastic particle schemes for large scale simulations relies on the ability to preserve a uniform distribution of particles in the whole physical domain. While simple particle split and merge algorithms have been considered previously, this study focuses on particle management based on a kernel density approach. The idea is to estimate the probability density of particles and subsequently draw independent samples from the estimated density. To cope with that, novel methods are devised in this work leading to efficient algorithms for density estimation and sampling. For the density inference, we devise a bandwidth with a bounded bias error. Furthermore, the sampling problem is reduced to drawing realizations from a normal distribution, augmented by stratified sampling. Thus, a convenient and efficient implementation of the proposed scheme is realized. Numerical studies using the devised method for direct simulation Monte-Carlo show encouraging performance.