gprHOG: Several Simple Improvements to the Histogram of Oriented Gradients Feature for Threat Detection in Ground-Penetrating Radar
share
The ground penetrating radar (GPR) is one of the most successful remote sensing modalities for buried threat detection (BTD). Substantial research has been devoted to the development of algorithms that automate BTD with GPR data, resulting in a large number of proposed algorithms. In this paper, we revisit the design of a recently and widely adopted algorithm based upon the popular Histogram of Oriented Gradients (HOG) feature. The HOG feature was originally developed in the computer vision community, and was subsequently applied to GPR-based BTD by Torrione et al., 2012. In this work, we propose several modifications to the HOG algorithm (i.e., as used in Torrione et al., 2012) with the aim of better adapting it to the unique characteristics of GPR data. Using a large GPR dataset, and several performance metrics, we show that each one of the proposed adaptations improves algorithm performance, sometimes by a substantial margin. We refer to the final algorithm, after all adaptations are included, as "gprHOG." This paper provides greater detail and experimental justification for the design of gprHOG, which was submitted to a recently published comparison of advanced algorithms for GPR-based BTD.
READ FULL TEXT