Computer vision techniques for target detection in ground penetrating radar data
Advances in ground penetrating radar (GPR) design, fabrication, and related signal processing have led to GPR becoming a powerful technology for the detection and identification of subsurface explosive threats. GPR's sensitivity to non-metallic disturbances in the subsurface makes it a highly complementary technology to electromagnetic induction (EMI) sensing, as GPR can sense very low-metal and non-metal threats that are difficult for EMI-based systems to detect. GPR based sensing typically requires the use of statistical learning algorithms to reduce false alarm rates caused by the sensor's increased sensitivity to non-metallic sub-surface objects. In recent years many feature extraction and statistical learning algorithms have been developed specifically for GPR processing with significant success. In separate development efforts, recent advances in the computer vision literature have resulted in feature descriptors and related classification algorithms that achieve robust results for a wide range of object-class identification problems in visual imagery. This work investigates the application of several of these techniques from the computer vision literature to target detection and classification in GPR data. In particular, in this work, histograms of oriented gradients (HOG) and Viola-Jones (VJ) features are extracted at locations of interest and used to classify object responses as explosive or non-explosive. Results indicate that HOG features perform quite well, outperforming the pre-screener and other specially-designed GPR object classification algorithms.
