Spectral Feature Extraction and Target Classification for Surface Penetrating Radar Sensing
Open Access
- Author:
- Tholl, Michael
- Graduate Program:
- Electrical Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- June 07, 2021
- Committee Members:
- Kultegin Aydin, Program Head/Chair
Ram Mohan Narayanan, Thesis Advisor/Co-Advisor
George Kesidis, Committee Member
Anthony Faust, Special Signatory - Keywords:
- Ground Penetrating Radar
Automatic Target Recognition
FDTD - Abstract:
- Landmine classification for a military use case is explored using simulated data generated by the open source Finite Difference Time Domain simulation software, gprMax. A variety of realistic mine and clutter targets are modelled in complex heterogeneous soil surfaces generated using realistic empirical models. Each of these models are simulated in gprMax, with the resulting data being processed using common GPR data processing techniques. Once the synthetic data have been pre-processed, Fourier-based techniques are used to extract spectral feature vectors representing the simulated target being tested. In addition to the aforementioned Fourier techniques, a procedure involving de-convolution of the received signal is proposed. This de-convolution procedure aims to better characterize the probed soil surface by accounting for, and removing, influence of the transmitted waveform. Following feature extraction, a variety of detection and/or classification based methods including a template matching based binary hypothesis test, Weighted K Nearest Neighbour Classifiers, and Support Vector Machines are proposed, and evaluated. The results show that template matching methods and trained classifiers can reliably discriminate between modelled discrete clutter objects and landmine targets in a variety of heterogeneous soil types, increasing the probability of detection without a corresponding increase in the probability of false alarms. This increase in the probability of detection without a corresponding increase in false alarms is due to the fact that the detection process occurs on a feature space extracted after a pre-screening process is conducted.