Real-time Adaptive Classification Environment using Rules (RACER)

"Robust Automatic Target Recognition (ATR) algorithms must be able to identify the conditions under which they are operating and tune their parameters to ensure targets are identified while retaining a low false-alarm rate. For example, the image signatureof a distant target is fundamentally different from the signature of a near target, so a single ATR algorithm with a single set of parameters will not perform optimally for both conditions. We propose an adaptive classification system called RACER(Real-time Adaptive Classification Environment using Rules). RACER uses meta-features, provided by a human operator or inferred from the scene, to identify the current operat-ing condition. Once the operating condition is identified, the on-line ATR'sclassifier and feature extraction parame-ters are tuned to maximize target recognition accuracy. The on-line tuning will be rule-based, with rules either speci-fied by an operator or automatically through reinforcement learning. A method for combining theoutput of multiple classifiers using Dempster-Shafer evidence combination will also be explored. Finally, a requirements study on hardware for a Phase II real-time implementation will be conducted. This includes assessing existing COTS recon-figurable FPGAboards, as well as compilers from C/C++ to the target FPGA. We see several potential applications of the proposed technology: 1) direct application of RACER to DoD ATR pro-grams, and 2) generalization of the classifier and learning algorithms to otherdomains, in particular the computer vision industry such as target learning & tracking of people using multi/hyperspectral imagery with applications to Homeland security."