Given the near-ubiquity of CCTV, there is significant ongoing research effort to apply image and video analysis methods together with machine learning techniques towards autonomous analysis of such data sources. However, traditional approaches to scene understanding remain dependent on training based on human annotations that need to be provided for every camera sensor. In this thesis, we propose an unusual event detection and classification approach which is applicable to real-world visual monitoring applications. The goal is to infer the usual behaviours in the scene and to judge the normality of the scene on the basis on the model created. The first requirement for the system is that it should not demand annotated data to train the system. Annotation of the data is a laborious task, and it is not feasible in practice to annotate video data for each camera as an initial stage of event detection. Furthermore, even obtaining training examples for the unusual event class is challenging due to the rarity of such events in video data. Another requirement for the system is online generation of results. In surveillance applications, it is essential to generate real-time results to allow a swift response by a security operator to prevent harmful consequences of unusual and antisocial events. The online learning capabilities also mean that the model can be continuously updated to accommodate natural changes in the environment. The third requirement for the system is the ability to run the process indefinitely. The mentioned requirements are necessary for real-world surveillance applications and the approaches that conform to these requirements need to be investigated. This thesis investigates unusual event detection methods that conform with real-world requirements and investigates the issue through theoretical and experimental study of machine learning and computer vision algorithms.