This report addresses the development of automated video-screening technology to assist security forces in protecting our homeland against terrorist threats. A prevailing threat is the covert placement of bombs inside crowded public facilities. Although video-surveillance systems are increasingly common, current systems cannot detect the placement of bombs. It is also unlikely that security personnel could detect a bomb or its placement by observing video from surveillance cameras. The problems lie in the large number of cameras required to monitor large areas, the limited number of security personnel employed to protect these areas, and the intense diligence required to effectively screen live video from even a single camera. Different from existing video-detection systems designed to operate in nearly static environments, we are developing technology to detect changes in the background of dynamic environments: environments where motion and human activities are persistent over long periods. Our goal is to quickly detect background changes, even if the background is visible to the camera less than 5 percent of the time and possibly never free from foreground activity. Our approach employs statistical scene models based on mixture densities. We hypothesized that the background component of the mixture has a small variance compared to foreground components. Experiments demonstrate this hypothesis is true under a wide variety of operating conditions. A major focus involved the development of robust background estimation techniques that exploit this property. We desire estimation algorithms that can rapidly produce accurate background estimates and detection algorithms that can reliably detect background changes with minimal nuisance alarms.
PDF
download
878987797
028-85220240
