Electroencephalography (EEG) and magnetoencephalography (MEG) are two non-invasive neuroimaging techniques aimed at localizing neural activity within the brain by measuring the electromagnetic field on the surface of the skull. These techniques are contributing to a better understanding of how the brain functions and are shedding light on the mechanisms of certain diseases, including epilepsy and Alzheimer’s disease. In this work we aim at improving spatio-temporal localization of interictal epileptic sources, which are known to exhibit a dynamic behaviour. We address the inverse problem using a Bayesian approach, which yields a probability distribution as a solution, thus allowing for a complete characterization of the uncertainty associated with the estimation. From a computational viewpoint we adopt algorithms belonging to the class of Particle Filters, that are especially advantageous when analyzing evolving dynamics within non-linear models; specifically, we adopt Rao-Blackwellized Particle Filters (RBPF), in which the posterior distribution is partly approximated analytically and partly sampled with Monte Carlo, exploiting a conditionally linear/Gaussian structure of the model. The main limitation of state-of-the-art RBPF implementations is that the reconstructed sources often travel considerable distances; however, this is in contrast with clinical evidence, which supports only limited movement. To address this issue we propose an updated dynamical model making use of a brain atlas and with the added constraint that sources are not allowed to move across different regions of the atlas. The proposed algorithm has been tested using different brain atlases, on both synthetic and experimental data; results indicate that the proposed method yields an improvement in the reconstruction accuracy.