Explosive hazards are one of the most deadly threats in modern conflicts. The U.S. Army is interested in a reliable way to detect these hazards at range. A promising way of accomplishing this task is using a<i> forward-looking ground-penetrating radar </i>(FLGPR) system. Recently, the Army has been testing a system that utilizes both L-band and X-band radar arrays on a vehicle mounted platform. Using data from this system, we sought to improve the performance of a <i>constant false-alarm-rate </i>(CFAR) prescreener through the use of a <i>deep belief network </i>(DBN). DBNs have also been shown to perform exceptionally well at generalized anomaly detection. They combine unsupervised pre-training with supervised fine-tuning to generate low-dimensional representations of high-dimensional input data. We seek to take advantage of these two properties by training a DBN on the features of the CFAR prescreener’s <i>false alarms </i>(FAs) and then use that DBN to separate FAs from true positives. Our analysis shows that this method improves the detection statistics significantly. By training the DBN on a combination of image features, we were able to significantly increase the probability of detection while maintaining a nominal number of false alarms per square meter. Our research shows that DBNs are a good candidate for improving detection rates in FLGPR systems.
This paper explores the effectiveness of an anomaly detection algorithm for downward-looking ground penetrating radar (GPR) and electromagnetic inductance (EMI) data. Threat detection with GPR is challenged by high responses to non-target/clutter objects, leading to a large number of false alarms (FAs), and since the responses of target and clutter signatures are so similar, classifier design is not trivial. We suggest a method based on a Run Packing (RP) algorithm to fuse GPR and EMI data into a composite confidence map to improve detection as measured by the area-under-ROC (NAUC) metric. We examine the value of a multiple kernel learning (MKL) support vector machine (SVM) classifier using image features such as histogram of oriented gradients (HOG), local binary patterns (LBP), and local statistics. Experimental results on government furnished data show that use of our proposed fusion and classification methods improves the NAUC when compared with the results from individual sensors and a single kernel SVM classifier.
Explosive hazard detection and remediation is a pertinent area of interest for the U.S. Army. There are many types of detection methods that the Army has or is currently investigating, including ground-penetrating radar, thermal and visible spectrum cameras, acoustic arrays, laser vibrometers, etc. Since standoff range is an important characteristic for sensor performance, forward-looking ground-penetrating radar has been investigated for some time. Recently, the Army has begun testing a forward-looking system that combines L-band and X-band radar arrays. Our work focuses on developing imaging and detection methods for this sensor-fused system. In this paper, we investigate approaches that fuse L-band radar and X-band radar for explosive hazard detection and false alarm rejection. We use multiple kernel learning with support vector machines as the classification method and histogram of gradients (HOG) and local statistics as the main feature descriptors. We also perform preliminary testing on a context aware approach for detection. Results on government furnished data show that our false alarm rejection method improves area-under-ROC by up to 158%.