Contextually Adaptive False Alarm Mitigation

In aerial images collected by wide-area sensors, targets of interest such as vehicles and dismounts are extremely small compared to the footprint of the image. The limited information that can be extracted from few pixels on target often means the targets of interest can be easily confused with other objects. Commonly used target detection systems in Intelligence, Surveillance and Reconnaissance (ISR) platforms can provide high detection rates but at the cost of high false alarm rates. Fortunately, potential objects of interest need not be viewed in isolation. Rather, when they are considered in the context of other information, many of the false alarms can be mitigated. In this effort, we propose a Machine Learning based approach to Contextual Rule Discovery for False Alarm Mitigation. A Learning Classifier System will be used to discover contextual rules of the form IF THEN , which can then be used to determine if a potential target flagged by an unreliable low level detector is indeed a true target. The Learning Classifier System must be trained to derive these rules, but we propose a method to conduct on-line automated training of the LCS without any human-in-the-loop. Contextual information is modeled through the automated generation of background class information. The system will be adaptive, in that background classes and contextual rules can be updated to deal with changes in environment during a mission. Scientific Systems Company, Inc. will be joined in this effort by Mercury Computer Systems and Dr. Robert Smith.