Neural network based video surveillance system

Video surveillance systems are usually composed of a network of active video sensors that continuously capture the scenes and present them to a human operator for analysis and event detection. Unfortunately human operators are often unable to monitor the video streams coming from a large number of video sensors. In this paper a semantic event detection system based on a neural classifier is presented to screen continuous video streams and detect relevant events, specifically for video surveillance. The goal of the proposed system is to automatically collect real-time information to improve the awareness of security personnel and decision makers. Our research is focused on the use of the "known scene rarr no alarm/unknown scene rarr alarm" paradigm, where the meaning of scene is related to spatial-temporal events, instead of the classical "frame difference" paradigm. The proposed system is able to detect mobile objects in the scene and to classify their movements (as allowed or disallowed) so as to raise an alarm whenever unacceptable movements are detected. This ability is supported also for video cameras mounted on a motorized pan scanner: experiments showed that the system is able to compensate the background changes due to the camera motion