Real-time Adaptive Classification Environment using Rules (RACER)

Automatic Target Recognition (ATR) algorithms must be able to identify the conditions under which they are operating and tune their parameters accordingly to ensure robust accuracy and low false-alarm rate. For example, the signature of a target imagedunder ideal weather conditions (overcast skies) is fundamentally different from the signature of a target imaged when raining. Therefore, a single ATR algorithm with a single set of parameters will not perform optimally for both conditions. In our Phase Ieffort, we developed a prototype algorithm capable of adapting a target detection algorithm to changing weather conditions. We propose to extend our approach to other operating conditions and to target recognition in Phase II. We call our adaptiveclassification system RACER (Real-time Adaptive Classification Environment using Rules). RACER uses meta-features, which are inferred from the scene, and current and past state information to identify the current operating conditions. Once identified, theon-line ATR's classifier and feature extraction parameters are tuned to maximize target recognition accuracy. In this effort we will also identity and categorize operating conditions that are problematic for infrared ATR. Finally, we will provide a workingprototype of RACER to Army Missile Command for an independent analysis using sequestered imagery. We see several potential applications of the proposed technology: 1) direct application of RACER to DoD ATR programs, and 2) generalization of the classifierand learning algorithms to other domains, in particular the computer vision industry such as target learning & tracking of people using infrared imagery with applications to homeland security.