In many surveillance missions information from a large number of interconnected sensors must be analysed in real time. When using visual sensors like CCTV cameras, it is not uncommon that an operator simultaneously has to survey the information from as many as fifty to a hundred cameras. It is obvious that the probability that the operator finds interesting observations is quite low when surveying information from that many cameras. In this paper we evaluate two different approaches for automatically detecting anomalies in data from visual surveillance sensors. Using the approaches suggested here the system can automatically direct the operator to the cameras where some possibly interesting activities take place. The approaches include creating structures for representing data, building "normal models" by filling the structures with data for the situation at hand, and finally detecting deviations in new data. One approach allows detections based on the incorporation of a priori knowledge about the situation combined with data-driven analysis. The other approach makes as few assumptions as possible about the situation at hand and builds almost entirely on data-driven analysis. The proposed approaches are evaluated off-line using real-world data and the results shows that the approaches can be used in real-time applications to support operators in civil and military surveillance applications.© (2008) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
2 October 2008