This work shows an unsupervised approach to model traffic flow and detect abnormal vehicle behaviors at intersections. In the first stage, the approach learns and discovers the different states of the system. The states are the result of coding and grouping the history motion of vehicles as long binary strings. In a second stage, using sequences of learned states, the authors build a stochastic graph model based on a Markovian approach. The authors label as an abnormal behavior when current motion pattern cannot recognize as any state of the system or a particular sequence of states cannot parse with the stochastic model. The authors tested our approach with several images sequences took from a vehicular intersection, where vehicular flow is continuously changing and traffic lights durations does not remains constant over day. Finally, the complexity and flexibility of the approach make it reliable to use in real time systems.