Comparison of feature extraction methods for landmine detection using Ground Penetrating Radar

Ground Penetrating Radar (GPR) senses dielectric discontinuities below the surface. Thus, it can detect low-metal and non-metal landmines. However, it detects not only landmines but also all objects under the ground and therefore, false alarm rates of GPR are very high. Powerful feature based algorithms are necessary to reduce false alarm rates and to distinguish landmine from clutter that causes false alarms. In this paper, Binary Robust Independent Elementary Features (BRIEF), Edge Histogram Descriptor (EHD), Histogram of Oriented Gradients (HOG), Scale Invariant Feature Transform (SIFT) and Speeded Up Robust Features (SURF) feature extraction methods are implemented to landmine detection problem. The methods are compared with extended data sets collected from different soil types by using surrogate landmines and other objects. Receiver Operating Characteristic (ROC) curves are calculated for comparison of methods and it is shown that the HOG outperforms other methods.