Remote concealed threat detection by novel classification algorithms applied to multi-polarimetric UWB radar

A method of effectively detecting remote concealed threats, particularly knives and guns, has been developed. This method uses multi-polarimetric ultra wide band active microwave radar to remotely scan a person under investigation. It has been shown that the radar signatures from such scans can be used to detect whether a person is carrying a concealed threat. A Principal Component Analysis (PCA) data reduction technique followed by a neural network (NN) is used to classify the information extracted from the radar signals. The technique combines the co, 45°, cross, and 135° polarized transceived radar signals into a single data set for classification. Illuminating the target with a range of polarizations, together with choosing a radar beam size commensurate with the targets in question, produces good discrimination between threat and non-threat items. Once collected, the data sets obtained are reduced via PCA, which significantly improves the correct classification rate at the NN stage and makes the technique more tolerant of variations in the threat objects orientation and better able to detect a wider range of threat types. Experimental results are presented which show that a detection rate of up to 80% for knives and guns can be achieved, with a false alarm rate as low as 4%.