Radars, as the name suggests, were traditionally used for radio detection and ranging. Nevertheless, advancement in technology has made it possible for using them as a sensor to identify or classify targets. The design of a reliable and robust (in term of noise and clutter) radar target recognition technique has long been the dream of many researchers.
Although target recognition based on Singularity Expansion Method (SEM) has been studied extensively for the last few decades, the feasibility of using the Complex Natural Resonances (CNRs) or poles to classify targets is further investigated in this thesis. CNRs are chosen for representing the target in this thesis as they are theoretically independent of the aspect angle between the radar and the target, and they form a minimal set of parameters by which the target can be identified thus assisting the classification problem.
This thesis focus on designing a reliable and robust resonance based target discrimination process by breaking down the resonance based target identification process into two parts. The first part involves addressing the computational issue of the poles extraction scheme and extracting the dominant poles of the target. In a resonance based target identification scheme, it is important to extract the contributing CNRs of the target and store them into a target feature reference library. Incorrectly determining the dominant poles of the target could result in a false alarm during the target identification step. The second part consists of designing a classifier that is capable of identifying the correct target from a set of targets in the presence of noise.
Matrix Pencil Method (MPM) is utilized in this thesis to extract the CNRs of the target. MPM is chosen due to its low sensitivity to background noise and its computational ease and efficiency. However, spurious poles as a result from overestimating the modal order of the system could deteriorate the performance of the MPM. A method has been proposed to overcome this issue and improve the performance of the MPM.
Two methods have been proposed in this thesis to extract the dominant poles of the target. The first novel approach uses Principle Component Analysis (PCA) to fuse the backscattering signatures of the target from either multiple incident aspects or multiple incident polarizations. The significant poles of the target could then be extracted from the fused signature. This method has the advantage of reducing the noise of the target’s signatures and thereby aiding the dominant CNRs extraction process. The second novel approach utilizes the energy of each of the extracted resonances of the target to identify the contributing poles of the target.
After extracting the significant CNRs of the target, the next step is to design a reliable and robust resonance based target identification technique. Two robust resonance based target identification schemes have been introduced in this thesis. The first unique method utilizes a standard tool in statistical analysis for target discrimination. Canonical Correlation Analysis (CCA), which has previously been used in economics and medical studies, has been proposed in this thesis for target identification. Numerical results show that this technique is comparable to the Generalized Likelihood Ratio Test (GLRT) in the presence of white Gaussian noise.
One of the issues related to resonance based target classifier is that it requires the commencement of the late time period for the unknown target response to be determined accurately, in order to avoid false alarms during the target classification process. For Automatic Target Recognition (ATR) applications, usually such information is not known a priori. In view of this problem, a modified GLRT technique that utilizes the Time-Frequency Analysis (TFA) is introduced in the second novel approach. The improved GLRT method does not require prior knowledge of the beginning of the late time period for the transient response of the unknown target. Simulation results show that the modified GLRT method is comparable to the original GLRT technique when the commencement of the late time period for the unknown target response is correctly determined and outperforms the original GLRT technique when the commencement of the late time period for the unknown target response is incorrectly determined.