Supervised learning methods for hand-held ground penetrating radar
Metadata[+] Show full item record
[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] Landmines and other buried explosives endanger military and civilian lives around the world. The detection of these types of subsurface objects using ground penetrating radar (GPR) is a difficult problem, partially due to variation in the size, content, and burial depth of different targets. Hand-held object detection systems pose yet another challenge, due to the inescapable human operator effect. In this thesis, computational methods for detecting dangerous targets in GPR data are explored. First, an anomaly detection algorithm which can act as a prescreener is described. Then, two feature extraction methods are implemented in order to extract relevant features from the GPR data at each prescreener alarm location. The first feature set describes hyperbolic edge features in the data by applying 36 log-Gabor filters to each B-scan and summing the energy over the filtered GPR data. The second feature extraction method collects local texture information from each B-scan by computing local binary patterns (LBP). Support vector machines (SVM) are trained to perform target vs. non-target classification using the different feature sets. Confidence values are then extracted from the trained SVMs and coupled with the prescreener confidence values in order to improve the overall detection performance.