Group detection represents an emerging Computer Vision research topic motivated by the increasing interest in modelling the social behaviour of people. This paper presents an unsupervised method for group detection which is based on an online inference pro-cess over Dirichlet Process Mixture Models. Formally, groups are modelled as compon-ents of an infinite mixture and individuals are seen as observations generated from them. The proposed sequential variational framework allows to perform inference in real-time, while social constraints based on proxemics rules ensure the production of proper group hypotheses consistent with human perception. The results obtained on several data-sets compare favourably with state-of-the-art approaches, setting the best performance in some of them.