Single sensor processing and sensor fusion of GPR and EMI data for landmine detection

In our previous work, we have shown theoretically that a model-based Bayesian approach to the detection of landmines affords significant performance gains over standard thresholding techniques. These performance gains hold for both time- and frequency-domain electromagnetic induction (EMI) sensors. Our methodology merges physical models of the evoked target response with a probabilistic description of the clutter. Under a specific set of assumptions, our technique provides both an optimal detection algorithm and performance evaluation measures expressed as probability of detection (Pd) and probability of false alarm (Pfa). This approach also provides a formal framework for incorporating target and/or environmental uncertainties into the processing algorithms. The significant performance improvements observed theoretically have been verified on both time-domain and frequency-domain EMI data collected in the field. In this paper, we review our previous theoretical work, and we use actual data collected in the field to illustrate the improvement obtained by appropriately accounting for environmental uncertainties. We present new results in which a suboptimal processor provides nearly identical performance to that of the optimal processor but with much greater computational efficiency. We also present results that indicate that such an approach can be applied successfully to ground penetrating radar (GPR) data. Specifically, we consider data taken by the BRTRC/Wichmann system. In addition to processing the data from each type of sensor individually, we discuss an approach where the data is fused using Bayesian techniques. The performance of each of these sensors individually, as well as the combination of sensors, will be discussed.