Study On Imaging Algorithm Based On Support Vector Machines For Through-the-wall Radars

The capabilities of electromagnetic waves to penetrate through nonmetallic materials provide a noninvasive opportunity to detect, localize, image and track hidden objects behind the barriers in ultrawideband through-the-wall radar(TWR) which is with a wide range of civilian and military applications. In the practical application, two key problems in TWR are ambiguities of wall parameters and real-time imaging. In order to solve them simultaneously, the studies based on the simulations of through-the-wall scenario using finite difference time domain(FDTD) are carried out in this dissertation, and the main content and achievements are as follows:Firstly, in oder to solve the recognize problem which most imaging algorithms can not realize, the near-field radar imaging models(with and without walls) are simulated, and a method based on support vector machine which can recognize the shape of the targets is proposed. The technology can recognite the shape of the targets effectively. The electromagnetic characteristics of the targets and the background are different(the classification label of every pixel point is different), and the amplitude of every pixel point is obtained by the time-domain algorithm and the frequency-domain algorithm, so the nonlinear relationship between the backscattered data of the targets and the classification label is established. The approximate expression of the relationship is obtained after the trainning processing of support vector machine, then the trainning model is used to estimate the classification label of every pixel point. The simulation results show that the process is helpful for increasing the classification accuracy of the targets and recogniting the shape of the targets effectively, and the estimated method only need no more than one second.Secondly, a two step imaging procedure in which firstly the wall parameters(relative permittivity, thick) are estimated used SVM and then the targets behind the wall is located and focused used PSM algorithm is presented. The procedure can satisfiy real-time imaging under unknown wall characteristics simultaneously. If the wall parameters are unknown, the estimated problem is converted to the establishment and use of a mapping between backscattered data and the wall parameters, then the nonlinear mapping is obtained after SVM training process. Because the target is unchanged If the SVM model is established, the influence of the change of the targets and the length of the wall、sampling interval、noise on the estimated wall parameters is mainly discussed, the results verify the feasibility and the validity of the approach based on SVM, and the approach has the advantages of high precision and low computational time. Fourther more, the procedure only need few tens of seconds to image with unknown wall parameters.Thirdly, a estimated method based on least squares support vector machines(LS-SVM) which can estimate the wall parameters(relative permittivity, thick, conductivity) is proposed. The entropy which is the yardstick of the imaging quality varifies that the estimated results are good or bad. The results are compared with that of the SVM-based approach. The technology can estimate the wall parameters sucessfully. If the wall parameters are estimated by LS-SVM, the estimated problem is converted to the establishment and use of a mapping between backscattered data and the wall parameters, then the nonlinear mapping is obtained after LS-SVM training process. Because the target is unchanged if the LS-SVM model is established, the influence of the change of the targets and the noise on the estimated wall parameters is mainly discussed, the estimated results show that the estimated parameters are affected by external factors greatly, and the change of the length of the wall, the sampling interval will nullify the LSSVM-based approach. Compared to the SVM-based approach, the LSSVM-based approach has higher precision and weaker generalization. If the practical application is considered, the SVM-based approach is more practical.Lastly, a real-time location method based on LS-SVM and SVM in the presence of wall ambiguities is proposed. The method avoids the wall effect and solves the real-time location problem under unknown wall parameters sucessfully. The location problem is converted to the establishment and use of a mapping between backscattered data and the location of the target, then the nonlinearity and the propagation effects caused by walls are included in the mapping that can be regressed after LS-SVM and SVM training process. The feasibility, validity and robustness of two methods are verified qualitatively and quantificationally. The simulated results demonstrate that the estimating process only need no more than one second after trainning process is completed, and the LSSVM-based approach has higher precision and more convenient than the SVM-based approach.