Design of multiple frequency continuous wave radar hardware and micro-Doppler based detection and classification algorithms
MetadataShow full item record
Micro-Doppler is defined as scattering produced by non-rigid-body motion. This dissertation involves the design of a multiple frequency continuous wave (MFCW) radar for micro-Doppler research and detection and classification algorithm design. First, sensor hardware is developed and tested. Various design tradeoffs are considered, with the application of micro-Doppler based detection and classification in mind. A diverse database of MFCW radar micro-Doppler signatures was collected for this dissertation. The micro-Doppler signature database includes experimental data from human, vehicle, and animal targets. Signatures are acquired from targets with varying ranges, velocities, approach angles, and postures. The database is analyzed for micro-Doppler content with a focus on its application to target classification. Joint time-frequency detection algorithms are developed to improve detection performance by exploiting noise-spreading and the micro-Doppler phenomenon. Following detection algorithm development, this dissertation covers the design of micro- Doppler feature extraction, feature selection, and classification algorithms. Feature selection is performed automatically via a Fisher score initialized sequential backward selection algorithm. Classification is performed using two distinct approaches: a generative statistical classification algorithm based on Gaussian mixture models (GMMs) and a discriminative statistical classification algorithm based on support vector machines (SVMs). Classifier performance is analyzed in detail on a micro-Doppler signature database acquired over a three-year period. Both the SVM and GMM classifiers perform well on the radar target classification task (high accuracy, low nuisance alarm probability, high F-measure, etc.). The performance of both classifiers is remarkably similar, and neither algorithm dominates the other in any performance metric when using the chosen feature set. (However, the difference between SVM and GMM classification accuracy becomes statistically significant when many redundant features are present in the feature set.) The accuracy of both classifiers is shown to vary as a function of approach angle, which physically corresponds to the angular dependence of micro-Doppler. The results suggest that overall classifier performance is more sensitive to feature selection than classifier selection (with GMM being more sensitive to redundant features than SVM). Both classifiers are robust enough to handle human targets attempting to evade detection by either army crawling or hands-and-knees crawling.