SBIR-STTR Award

The goal of this project is to develop a neural-network based system for air-to-air aircraft identification from High Resolution Range (HRR) radar target return signatures. Training and testing of this system will be carried out using appropriate data sets from the large data base of HRR-radar-acquired signatures of known targets residing at the Wright Patterson Air Force Base. In performing target identification, the system will optimally nonlinearly map the observed target signature x into a low-dimensional feature vector z, and then classify z according to an appropriate statistical decision rule like Neyman-Pearson's. For this purpose, the statistics of z will be acquired from the training data by nonparametric methods. In order to increase robustness and accuracy, the system will be implemented in the form of a hierarchical classifier. In such an implementation, provision will be made for bringing information from other sources to reduce uncertainty through data fusion. Preliminary results on a limited data set look encouraging.

Keywords:
RADAR INTELLIGENT SYSTEM IDENTIFICATION SIGNAL PROCESSING PATTERN RECOGNITION

This project addresses the Non-Cooperative Target Identification (NCTI) problem which is important for the Air Force's Automatic Target Recognition (ATR) project at the Wright Laboratory. In the past the ATR project produced a system for air-to-air aircraft identification by classifying High Resolution Range (HRR) radar return signatures (measured signatures) using a classifier with parameters generated either from synthetic signatures or form a combination of synthetic and measured signatures. The test performance using current methods, however, has been unsatisfactory.The goal of this project is to further methods, developed in Phase I SBIR, to advance the ATR objective by improving classifier parameters using measured signatures, synthetic signatures, and geometric models of the target. Speicifically our approach attempts to correct synthetic classifier templates by adjusting the target modleing parameters based on available measured data. Preliminary results using adjusted synthetic signatures show an improved matching between the synthetic data and the measured data. Altough the results are preliminary, they show potential for generalization. In addition, this SBIR will continue to explore the possiblity of using a neural-network-based classifier to enhance the NCTI problem.