The Research Of Feature Extraction Method Based On Radar Echo Data

Radar target recognition, which is an important part of modern radar system, hasbeen widely used in military and civilian fields. However, the feature extraction is thekey step of the target recognition system. It is important to choose what kind of thefeature which can character the target. Based on the measured data of radar echo, thisthesis makes a deep study on feature extraction of the radar target recognition system.Firstly, it introduces the scattering center model of radar target and range profilebased on scattering center model, and discusses the range profile characteristics of radartarget. Because high resolution range profile of radar target can provide good targetstructure and shape information, many methods of feature extraction, which is based onhigh resolution range profile, are studied intensively in this thesis. Then the measureddata of radar echo is described, which forms the basis for the following study methodsof radar target feature extraction.Secondly, in order to solve the translation sensitivity of range profile, three kindsof translation invariant features are deeply introduced, power spectrum, autoregressioncoefficient and central moments. A computationally efficient method is proposed forfeature extraction based on even rank central moments of high resolution range profile.The experimental results based on real radar data show that the proposed methodachieves good robustness and accuracy with its reduced storage of the template featurevectors and calculation of the test sample identification. Then, with a study onfeature-level fusion, power spectrum and autoregression coefficient are fused byprinciple component analysis method. Experimental results show that the PCA fusioncan reduce the number of dimensions in a sufficient number of valid information,improving the efficiency of target identification.Finally, the original data of high resolution range profile is mapped tohigh-dimensional data space by using nonlinear kernel function method. Analyzinghigh-dimensional data by principle component analysis extract the easier kernelcomponent for identification. Extracting nonlinear feature can effectively compensatefor the deficiencies of traditional feature. Experimental results show the nonlinearkernel feature extraction algorithm is effective and feasible.