Target detection and classification using seismic signal processing in unattended ground sensor systems

This thesis studies the problem of target detection and classification in Unat-tended Ground Sensor (UGS) systems. One of the most challenging problems faced by target identification process is the design of a robust feature vector which is sta-ble and specific to a certain type of vehicle. UGS systems have been used to detect and classify a variety of vehicles. In these systems, acoustic and seismic signals are the most popularly used resources. This thesis studies recent development of target detection and classification techniques using seismic signals. Based on these studies, a new feature extraction algorithm. Spectral Statistics and Wavelet Coef-ficients Characterization (SSWCC), is proposed. This algorithm obtains a robust feature vector extracted from the spectrum, the power spectral density (BSD) and the wavelet coefficients of the signals. Shape statistics is used in both spectral and PSD analysis. These features not only describe the frequency distribution in the spectrum and PSD, but also shows the closeness of the magnitude of spectrum to the normal distribution. Furthermore, the wavelet coefficients are calculated to present the signal in the time-frequency domain. The energy and the distribution of the wavelet coefficients are used in feature extraction as well. After the features are obtained, principal component analysis (PGA) is used to reduce the dimension of the features and optimize the feature vector. Minimum-distance classifier and k-nearest neighbor (kNN) classifier are used to carry out the classification. Experimental results show that SSWCC provides a robust feature set for target identification. The overall performance level can reach as high as 90%.