Bayesian sensor fusion for multi-platform landmines detection

Abstract

This paper presents a novel sensor fusion model able to combine data from Ground Penetrating Radar (GPR) and Metal Detectors (MD) in order to classify landmines in a scanned floor area. Currently, no sensor detects landmines directly: a metal detector detects landmines' metal content and a ground penetrating radar detects dielectric discontinuities in the soil, that may be generated by a buried landmine. Fusing the information from different types of such sensors would improve the Probability of Detecting (PoD) landmines and decrease the rate of False Alarms (FAR). The current work describes a Bayesian decision level fusion which was found to decrease the FAR and improve the PoD when compared with data level and feature level fusion approaches. The classifier was tested using different sensors attached to different mobile platforms and after geo-referencing the acquired data, and training the proposed fusion classifier over a large experimental data set containing landmines and other objects, significant improvements both in the PoD and FAR were observed. The presented results are based on data acquired with an IDS GPR array and two Vallon MD arrays, in DOVO military facilities, near Leuven, Belgium, during July 2014.