Résumé : In this work we present a system to extract in an unsu- pervised manner the main activities that can be observed by a camera monitoring a scene on the long-term and with the ultimate aim to discover abnormal events. To allow for semantically interpretable results, the activities are characterised by referring them to contextual elements of the observed scene. By contextual elements, we understand activity areas such as building entrances, people meeting areas, road areas,...etc. The system thus stars in a first step by the unsupervised learning of the main activity areas of the scene. We employ trajectory-based analysis of mobiles in the video to discover the points of entry and exit of mobiles appearing in the scene and ultimately deduce the different areas of activity. In a second step, mobile objects are then characterised in relation to the learned activity areas. Two kind of behaviours can then be defined either 'staying in a given activity zone' or 'transferring from an activity zone to another' or a sequence of the previous two behaviours. In a third step we employ a high-level clustering algorithm to group mobiles according to their behaviours and discover both, frequent/normal behaviours and unusual/abnormal events. We employ soft computing techniques in the first two steps to handle the uncertainty inherently present in low-level trajectory data. The high-level clustering algorithm is based on relational analysis methodology. We have applied our system in two domains, the monitoring of activities in the hall entrance of an underground station and to traffic control by monitoring a bus reserved street lane. We show that our current results are encouraging.