We propose an event learning approach for video, based on concept formation models. This approach incrementally learns on-line a hierarchy of states and event by aggregating the attribute values of tracked objects in the scene. The model can aggregate both numerical and symbolic values. The utilisation of symbolic attributes gives high flexibility to the approach. The approach also proposes the integration of attributes as a doublet value-reliability, for considering the effect in the event learning process of the incertainty inherited from previous phases of the video analysis process. Simultaneously, the approach recognises the states and events of the tracked objects, giving a multi-level description the object situation. The approach has been evaluated for an elderly care application and a rat behaviour analysis application. The results show that the approach is capable of learning and recognising meaningful events occurring in the scene, and to build a rich model of the objects behaviour. The results also show that the approach can give a description of the activities of a person (e.g. approaching to a table, crouching), and to detect abnormal events based on the frequency of occurrence.