Duke University Department of Electrical and Computer Engineering

Physics Model-based Signal Processing of Ground Penetrating Radar 
for Subsurface Object Detection and Discrimination

Vivek Munshi, Senior Pratt Engineering Undergraduate Fellow

Major: Electrical and Computer Engineering/Computer Science

Advisor: Dr. Leslie Collins

Ground penetrating radar (GPR) has been proposed as an alternative sensor to classical electromagnetic induction techniques for the problem of subsurface object detection. Traditional GPR sensors aim for the detection of dielectric discontinuities below the surface. As with most fielded sensors, the processing that is performed is essentially anomaly detection, thereby resulting in excessively large false alarm rates. These large false alarm rates prove to be very costly with respect to both the time lost and expensive costs of digging where there is no object of interest.

One of the major goals of this project is the development and application of an electromagnetic model of GPR to predict measured responses from buried landmines. The second major goal is the development of signal processing algorithms to discriminate landmines from clutter. The data being analyzed is that which was collected with the Wichmann-NIITEK radar. The Wichmann-NIITEK antenna provides a high frequency radar signal with very low noise levels following the ground reflection. The algorithm was tested on data collected with this radar in the calibration area and blind grid at the Hand Held Landmine Detector Performance Baselining test site developed by the Joint UXO Coordination Office (JUXOCO).

Two time-domain algorithms and one frequency-domain algorithm for subsurface object detection have been proposed up to this point. The time-domain algorithms attempt to remove the undesired ground response from measured data, which effectively enhances signals from subsurface objects. The frequency-domain algorithm detects the energy of a specific range of the magnitude response, disregarding frequency content from the ground response. The results obtained so far indicate that the time-domain and frequency-domain algorithms should be used in parallel for optimal object detection.

Utilizing the detection methods described above, we can determine “areas of interest” from the measured data. Given these areas of interest, we explored algorithms to discriminate landmines from clutter. Certain features can be extracted from the measured response with the aid of Matching Pursuits (Mallat 1993). We inputted and trained this feature set into automated learning tools, such as neural networks and support vector machines, in order to differentiate between responses from landmines and those from clutter. The algorithms employing support vector machines correctly classified, with high probability, landmine signals from clutter signals. The detection and discrimination algorithms are now being tested on blind lanes to better gauge performance. These algorithms are also being run on data collected from a larger sensor on different terrains such as gravel and dirt to determine the effect of surface material on detection.

The full final report is available here in Adobe Acrobat Reader (pdf) format.