Radar Target Recognition

Radar target recognition(RTR) is very valuable for military application. The performance of RTR can be improved greatly by identity fusion. This dissertation deals with the RTR using data level and decision level fusion.On data level, it is firstly explored fusing signature data from sparse-band collocated radars to obtain wider band target frequency response. The most difficult problem which restricts the fusion process is the lack of mutual coherence between the various radar subbands. This dissertation analyzes the causations of noncoherence, develops the concrete mathematical relation between sparse-band radar signatures and presents a method to compensate the lack of mutual coherence. The simulation results show that this method is effective. Then this dissertation investigates the methods of radar target recognition based on one dimensional range profiles and one dimensional scattering centers which are obtained from target wideband frequency responses. Two methods are presented, one is based on the normalized central moments of one dimensional range profiles, the other is based on one dimensional scattering centers matching. Applying the two methods to five target scale models data measured at outfield achieves good recognition results.On decision level, this dissertation deals with fuzzy methods and neural network methods. In order to reduce the disadvantageous influence of decision profiles' scattering on fusion recognition, a decision level fusion scheme for target recognition based on K-nearest neighbor decision profiles is presented. The definitions of various fuzzy integrals are introduced. And their intuitionistic meanings for decision level fusion are interpreted. The core problem in using fuzzy integral for decision level fusion is to determine the fuzzy densities. A method of determining fuzzy densities adaptively is presented, which uses the apriori static information of the training samples and the dynamic information contained in each sensor's decision. If we concatenate the recognition results of each sensors and consider it as a joint feature vector of target, then decision level fusion for target recognition can be considered as a conventional pattern recognition problem. This dissertation uses multilayer perceptron to classify the joint feature vector. Anew neural network algorithm for decision level fusion is also presented in this dissertation. The architecture of this network is novel. It is the thresholds, not the conjunction weights, which are modified, when the network is being trained. The simulation experiment results show that these methods are effective.