This thesis investigates detection and classification issues when dealing with seismic signals and represents a first step in the direction of automated detection and classification of mine-like signals obtained using a seismic approach. A computationally cheap detection scheme that utilizes a combination of a simple combination of a short- term energy and zero-crossing detector is implemented and tested on five different classes of targets, resulting in a 100% detection rate for all non-natural targets and 33% detection rate of mine sized rock buried in sand. Three feature extraction methods are evaluated for their possible use in a Gaussian Mixture Model classifier: higher order moments, pole extraction from impulse response modeling using the Steiglitz-McBride iteration, and Radial Basis Function Modeling of data. These methods demonstrate promising results for use in a classifier. However, only a very limited number of data trials per class was available in this work, and the proposed set-up needs to be further validated with additional data.
展开▼
