Long-term surveillance performance improvement using exploitation memory

We describe an experiment to quantify the exploitation performance benefits of acquired site-specific algorithm performance knowledge, applied to the problem of motion detection. Automated object detection algorithms typically suffer from high false alarm rates that can be distracting to an operator, in some cases rendering the algorithm useless. The experiment is designed to reduce the false alarm rate of a surveillance algorithm in scenarios that involve repeated observation of the same location using an algorithm as a cuer to an operator who accepts or rejects algorithm reports. The experiment involves using a wrapper that records the operator decisions, and modulates algorithm reports using acquired knowledge of high-false-alarm areas and high-true-detection areas. The wrapper also uses an a priori operator segmentation of the scene that marks some regions for which no algorithm reports will be passed on to the operator. The experiment demonstrates a 35% reduction in the time required for an analyst to find vehicles in a search scenario, without any negative impact on accuracy. (C) 2007 Society of Photo-Optical Instrumentation Engineers.