Feature study for high-range-resolution based automatic target recognition: Analysis and extraction

Techniques for automated moving target/object recognition are required in many military and civilian applications. This dissertation focuses on moving target Automatic Target Recognition (ATR) using High Range Resolution (HRR) radar, with a special emphasis on features study.; After briefly discussing the basics of HRR radar sensors, the objectives of this investigation are formally defined in Chapter 1. In Chapter 2, an extensive review of previous research into HRR ATR is provided. In Chapter 3 we investigate the utility of complex HRR signatures based on both theoretical analysis and experimental examinations. Chapters 4 and 5 are devoted to finding and extracting robust HRR signatures. In Chapter 4 we derive a physics-based HRR moving target model, and define the parameters of the model as a set of potential features. Additionally, two parameter estimation algorithms based on this model are developed. Subsequently these algorithms can serve as the feature-extraction algorithms for the HRR data described by our proposed model. However, the features defined and extracted in Chapter 4 are representational features, and we thus cannot guarantee that they represent distinguishing information between targets. Thus, Chapter 5 approaches the feature extraction problem differently, and a new nonlinear feature extraction algorithm, named Kernel-based nonlinear Feature Extraction (KFE) algorithm, is proposed. This algorithm extends conventional linear scatter matrix-based feature extraction algorithms to the nonlinear domain via a technique referred to as the “kernel trick”. Both theoretical proofs and experimental tests demonstrate that high performance features can be extracted using this KFE algorithm. Because SVM, an emerging technique in pattern recognition, is extensively involved in Chapter 5, a brief introduction to it is provided in Appendix B.; The main contributions of this research are (1) a deeper understanding of the properties of complex HRR signatures, (2) a physics-based HRR moving target model, (3) a set of physical feature extraction algorithms based on our proposed HRR models, and (4) a new category of nonlinear algorithms which can be used to extract discriminant features.