In recent years, advancements in the field of data acquisition from satellite navigation systems and automatic identification systems (AIS) has led to rapid increase of data in the domain of maritime surveillance. However, given the complexity and great volume of trajectory data, traditional approaches for analysing and exploring such information poses several limitations. For that reason, an automated method for identifying common modes of behaviour in distinct trajectory segments could provide significant benefits, particularly in the problem of event recognition and trajectory prediction. This thesis approaches this problem as an imitation learning task and explores the use of directed Info-GAIL algorithm, along with historical AIS data and its contextual features, in order to learn subtask policies from unsegmented demonstrations of vessel trajectories.