Using Pattern Recognition with HOG to Automatically Detect Reflection Hyperbolas in Ground penetrating Radar Data

Ground penetrating radar (GPR) is used for nondestructive examination of utility present underground such as cables, pipes, and landmines. Cables and pipes make hyperbolic shape signatures in GPR radargrams. Automatic detection of hyperbolic signature narrow down the region of interest and therefore results in reduced data set for localization of the buried pipe. In this paper, a machine learning approach is used to detect hyperbolic signatures using a support vector machine (SVM) with the histogram of oriented gradient features (HOG). For this purpose, Voila Jones algorithm is used in Matlab to train a classifier with HOG features instead of haar features. The HOG features have not been used with SVM in the literature for hyperbolic signature detection in the ground penetrating radar (GPR) radargrams. In this paper, it is shown that HOG feature based classifier achieve a high detection rate of 0.758 with a low false positive rate of 0.394. The detection algorithm is tested on both real GPR data and synthetic GPR data. Synthetic GPR data is created on an open source software gprMax.