An unsupervised multi-feature framework for landmine detection

A multi-feature framework for the detection of antipersonnel landmines with Ground Penetrating Radar (GPR) is suggested. The features result from independently acquired and processed GPR measurements. The initial detection in the confidence maps is made independently after which these detection coordinates are co-located. The marginal feature distributions are normalized via Johnson;s transform prior to the process of their fusion. A Maximum Likelihood based classifier is used as a fusion operator. The operator takes a quadratic form due to the enforced normality of the feature distributions. The framework trains the classifier using secondary data acquired at an open site. The framework;s performance is illustrated using the data acquired over a specifically designed test-site.