Hunter, Andrew and Owens, Jonathan and Carpenter, M. (2003) A Neural System for Automated CCTV Surveillance. In: IEE Symposium Intelligent Distributed Surveillance Systems, 26 February 2003, London.
|Item Type:||Conference or Workshop Item (Paper)|
|Divisions:||College of Science > School of Computer Science|
|Abstract:||This paper overviews a new system, the “Owens Tracker,” for automated identification of suspicious pedestrian activity in a car-park. Centralized CCTV systems relay multiple video streams to a central point for monitoring by an operator. The operator receives a continuous stream of information, mostly related to normal activity, making it difficult to maintain concentration at a sufficiently high level. While it is difficult to place quantitative boundaries on the number of scenes and time period over which effective monitoring can be performed, Wallace and Diffley  give some guidance, based on empirical and anecdotal evidence, suggesting that the number of cameras monitored by an operator be no greater than 16, and that the period of effective monitoring may be as low as 30 minutes before recuperation is required. An intelligent video surveillance system should therefore act as a filter, censuring inactive scenes and scenes showing normal activity. By presenting the operator only with unusual activity his/her attention is effectively focussed, and the ratio of cameras to operators can be increased. The Owens Tracker learns to recognize environmentspecific normal behaviour, and refers sequences of unusual behaviour for operator attention. The system was developed using standard low-resolution CCTV cameras operating in the car-parks of Doxford Park Industrial Estate (Sunderland, Tyne and Wear), and targets unusual pedestrian behaviour. The modus operandi of the system is to highlight excursions from a learned model of normal behaviour in the monitored scene. The system tracks objects and extracts their centroids; behaviour is defined as the trajectory traced by an object centroid; normality as the trajectories typically encountered in the scene. The essential stages in the system are: segmentation of objects of interest; disambiguation and tracking of multiple contacts, including the handling of occlusion and noise, and successful tracking of objects that “merge” during motion; identification of unusual trajectories. These three stages are discussed in more detail in the following sections, and the system performance is then evaluated.|
|Date Deposited:||19 Jun 2009 10:08|
Actions (login required)