Radar target classification based on micro-Doppler signature analysis

Automatic target recognition based on features obtained from micro-Doppler signatures has a great potential for surveillance as well as data collection applications. In order to guide the feature extraction, a physical understanding of the micro-Doppler phenomenon has been investigated by means of the electromagnetics theory. The electric field scattered by a rotating cylinder has been modelled by the Method Of Moments and a high matching has been observed with the measurements carried out with a Doppler radar. This provides insight into the response of more complex (rotating) targets and allows for the characterisation of the data acquisition setup. The Physical Optics approximation has been used to model the electric field scattered by the human body, whose motion is computed from MOCAP data. The model has been validated by comparing the simulated and measured micro-Doppler signatures induced by canonical movements of the leg and the arm. The target classification has been performed on 476 collected data distributed over seven classes, including human activities and vehicles. A sequential forward selection algorithm combined with Fisher scores is used to select seven features out of the 57 extracted ones. The performance of a Naive Bayes classifier and several Support Vector Machine classifiers is assessed using 10-fold cross-validation. All the classifiers reach an accuracy above 90 %. Hence, the classifier choice is shown to be less critical than the choice of the features.