Automatic target detection in GPR images using Histogram of Oriented Gradients (HOG)

Abstract

Ground Penetrating Radar (GPR) has proven itself to be one of the most popular and reliable geophysical device in subsurface investigation. However, human operators are required to manually interpret the GPR data. In a typical geophysical survey, collected GPR data sometimes can be enormously huge, causing issues such as time consuming and inaccuracy in results due to human errors. In this paper, we present an algorithm that automatically detects hyperbolic signatures in GPR data in B-scan model. This developed algorithm is able to mark potential regions that contain the reflections from target of buried objects. Histogram of Oriented Gradients (HOG) was initially developed to detect pedestrians, but it can be also well-adapted to detect particular shapes and objects. HOG descriptors are extracted from a set of training images and are trained using a linear SVM classifier. The main purpose of this algorithm is to narrow down the data into possible target reflection regions. After that, we implement Hough Transform to highlight the hyperbolic patterns in the reflection. The results shows that the developed system can perform target detection at an average of 93.75% detection rate for all four test sets.