Optimisation de la détection CFAR distribuée et estimation dans un clutter de distributions non gaussiennes.

Abstract

The objectives of this thesis deal with the domain of radar detection and estimation. The first problem is the study of the optimization problem of distributed CFAR (Constant False Alarm Rate) detection in a Pareto type I distribution environment. In this part, we analyze the performance of the distributed CFAR detectors GM-(Geometric Mean), OS- (Ordered Statistic), GO- (Greatest of) and SO-(Smallest of) CFAR in a Pareto clutter. In this case, we first obtain the approximate expressions for the PD for the GM-CFAR, OS-CFAR, GO-CFAR and SO-CFAR detectors. Due to the nonlinear property of this multidimensional system, we propose the use of an efficient optimization approach based on the BBO (Biogeography Based Optimization) algorithm to obtain the optimal scale factor of the local detectors. Each detector makes its own decision and sends it to the fusion center to obtain a binary global decision according to a preselected fusion rule. We examine the cases of three data fusion rules ""AND"", ""OR"" and ""MAJORITY"" at the fusion center. Through Monte-Carlo simulations, the detection performances of the detectors are evaluated for a homogeneous and heterogeneous Pareto clutter. For high-resolution radars, sea clutter modeling has shown that CG distributions are appropriate for describing these clutter returns. The second study addresses the problem of adaptive CFAR detection in a non-Gaussian clutter. We propose three new CFAR detectors in a non-coherent context, where the clutter follows a non-Gaussian distribution. Monte Carlo simulations have shown that the new detectors are robust for three CG models; namely the K distribution, the Compound Gaussian distribution with Inverse Gamma texture (GP) and the CIG distribution. The false alarm regulation was then examined in the presence of interfering targets. Finally, the performance of the three proposed algorithms were validated using real IPIX data The third objective concerns the CFAR detection in a log-normal clutter. The proposed algorithm is based on scale and power invariant distributions. This includes the choice of two functions called scale invariant function and secondary function of CRP (Clutter Range Profile). However, the existing CFAR algorithms exhibit remarkable CFAR losses due to the presence of outliers. In order to provide a modified decision rule with immunity against interfering targets, we propose an appropriate choice of these two functions. To do this, two functions based on WH and ordered statistics are proposed for a log-normal clutter. The dependence of the false alarm probability on interfering targets and the log-normal distributed clutter parameters are also investigated. From the simulated data, the log-t detector, GMOS-(Geometric Mean Ordered Statistic), TMOS-(Trimmed Mean Ordered Statistic), IE-(Inclusion / Exclusion) and WH-(Weber-Haykin)CFAR detectors are used for the purpose of comparison. The results obtained from synthetic data clearly indicate that a smaller CFAR loss is obtained by the proposed decision rule and outperforms the other detectors in the presence of multiple interfering targets. A lower CFAR loss is obtained by the proposed decision rule, in particular in the presence of strong secondary targets. Finally, the performance of the proposed algorithms were validated using real IPIX data The forth objective deals with the CFAR censored maximum likelihood detection in a Gamma distribution environment with a known shape parameter. In this study, we propose the (Censored Maximum Likelihood Estimate) CMLE -CFAR detector under the case of one censored sample. The decision rule of the proposed CMLE detector is given in terms of ML estimates of the scale parameter. Based on the Monte-Carlo simulations, the detection performances of the CML-CFAR detector are compared to the existing CA-, ML- and OS CFAR algorithms. In the presence of interfering targets, it is shown that there is an improvement in the probability of detection if the proposed CMLE-CFAR algorithm is used. The fifth problem in this thesis consists of developing the CMLE and Bayes methods to estimate the dispersion parameter of the Pearson population from censored samples. The proposed estimators cannot be obtained in closed forms in which the estimates are runed numerically after setting the desired number of censored data.