Résumé : The abundance of video surveillance systems has created a dire need for computational methods that can assist or even replace human operators. Research in this field, however, has yet to tackle an important real-world scenario: extremely crowded scenes. The excessive amount of people and their activities in extremely crowded scenes present unique challenges to motion-based video analysis. In this paper, we present a novel statistical framework for modeling the motion pattern behavior of extremely crowded scenes. We construct a rich yet compact representation of the local spatio-temporal motion patterns and model their temporal behaviors with a novel, distribution-based Hidden Markov Model (HMM), exploiting the underlying statistical characteristics of the scene. We demonstrate that, by capturing the steady-state behavior of a scene, we can naturally detect unusual events as unlikely motion pattern variations. The experiments show promising results in extremely crowded real-world scenes with complex activities that are hard for even human observers to analyze.