Equipe OBELIX, UBS/IRISA, Vannes AIS data gathers high frequency positions and speed data of world marine traffic, in order to prevent collisions between vessels. An important challenge is to develop efficient automatic procedures to detect abnormal behaviours within large streams of AIS data. As a first step to reach this objective, trajectory data must be clustered in homogeneous parts. The challenge here is that the AIS stream gathers heterogeneous data at two levels: 1. At the traffic's

more »

... l: Positions data comes from different types of vessels, therefore leading to many different trajectory patterns; 2. At the trajectory's level: It is known that movement patterns within a trajectory can be different depending on the individual's behaviour. We propose a global unsupervised approach that aims to perform clustering at these two levels such that, for an observed trajectory: • The trajectory is segmented in homogeneous movement patterns; • The trajectory is assigned to a cluster of trajectories sharing the same movement patterns. The unsupervised approach is performed using a dual hierachical dirichlet processes. In order to have a scalable estimation algorithm, we use stochastic variationnal inference. I'll present the model developped, as well as the inference method used. I'll show first results on a set of cargos trajectories.