This PDF file contains the front matter associated with SPIE Proceedings Volume 9072, including the Title Page, Copyright information, Table of Contents, Introduction, and the Conference Committee listing.

Learned representations have been shown to give hopeful results for solving a multitude of novel learning tasks, even though these tasks may be unknown when the model is being trained. A few notable examples include the techniques of topic models, deep belief networks, deep Boltzmann machines, and local discriminative Gaussians, all inspired by human learning. This self-learning of new concepts via rich generative models has emerged as a promising area of research in machine learning. Although there has been recent progress, existing computational models are still far from being able to represent, identify and learn the wide variety of possible patterns and struc- ture in real-world data. An important issue for further consideration is the use of unsupervised representations for novel underwater target recognition applications. This work will discuss and demonstrate the use of latent Dirichlet allocation and autoencoders for learning unsupervised representations of objects in sonar imagery. The objective is to make these representations more abstract and invariant to noise in the training distribution and improve performance.

In this paper, an automated change detection technique is presented that compares new and historical seafloor images created with sidescan synthetic aperture sonar (SAS) for changes occurring over time. The method consists of a four stage process: a coarse navigational alignment; fine-scale co-registration using the scale invariant feature transform (SIFT) algorithm to match features between overlapping images; sub-pixel co-registration to improves phase coherence; and finally, change detection utilizing canonical correlation analysis (CCA). The method was tested using data collected with a high-frequency SAS in a sandy shallow-water environment. By using precise co-registration tools and change detection algorithms, it is shown that the coherent nature of the SAS data can be exploited and utilized in this environment over time scales ranging from hours through several days.

The ability to autonomously sense and characterize underwater objects in situ is desirable in applications of unmanned underwater vehicles (UUVs). In this work, underwater object recognition was explored using a digital holographic system. Two experiments were performed in which several objects of varying size, shape, and material were submerged in a 43,000 gallon test tank. Holograms were collected from each object at multiple distances and orientations, with the imager located either outside the tank (looking through a porthole) or submerged (looking downward). The resultant imagery from these holograms was preprocessed to improve dynamic range, mitigate speckle, and segment out the image of the object. A collection of feature descriptors were then extracted from the imagery to characterize various object properties (e.g., shape, reflectivity, texture). The features extracted from images of multiple objects, collected at different imaging geometries, were then used to train statistical models for object recognition tasks. The resulting classification models were used to perform object classification as well as estimation of various parameters of the imaging geometry. This information can then be used to inform the design of autonomous sensing algorithms for UUVs employing holographic imagers.

Recent research has shown that synthetic aperture acoustic (SAA) imaging may be useful for object identification. The goal of this work is to use SAA information to detect and identify four types of objects: jagged rocks, river rocks, small concave capped cylinders, and large concave capped cylinders. More specifically, we examine the use of frequency domain features extracted from the SAA images. We utilize Support Vector Machines (SVMs) for target detection, where an SVM is trained on target and non-target (background) examples for each target type. Assuming perfect target detection, we then compare multivariate Gaussian models for target identification. Experimental results show that SAA-based frequency domain features are able to detect and identify the four types of objects.

A recently validated technique for buried target detection relies on applying an acoustic stimulus signal to a patch of earth and then measuring its seismic (vibrational) response using a laser Doppler vibrometer (LDV). Target detection in this modality often relies on estimating the acoustic-to-seismic coupling ratio (A/S ratio) of the ground, which is altered by the presence of a buried target. For this study, LDV measurements were collected over patches of earth under varying environmental conditions using a known stimulus. These observations are then used to estimate the performance of several methods to discriminate between target and non-target patches. The first part of the study compares the performance of human observers against a set of established seismo-acoustic features from the literature. The simple features are based on previous studies where statistics on the Fourier transform of the acoustic-to-seismic transfer function estimate are measured. The human observers generally offered much better detection performance than any established feature. One weakness of the Fourier features is their inability to utilize local spatiotemporal target cues. To address these weaknesses, a novel automatic detection algorithm is proposed which uses a multi-scale blob detector to identify suspicious regions in time and space. These suspicious spatiotemporal locations are then clustered and assigned a decision statistic based on the confidence and number of cluster members. This method is shown to improve performance over the established Fourier statistics, resulting in performance much closer to the human observers.

Typical metal detectors work at very low to low frequencies. In this paper, a metal/anomaly detector design that operates in the high to very high frequency range is presented. This design uses a high-Q tuned loop antenna for metal/anomaly detection. By measuring the return loss or voltage standing wave ratio a frequency notch can be detected. Tuning to the optimal location of the notch can be accomplished by monitoring the phase response. This phase monitoring technique can be used to ground balance the detector. As a metal object is moved along the longitudinal axis of the loop antenna a substantial shift in the frequency of the notch is detected. For metal targets, the frequency shift is positive, and for ferrite and other targets, the frequency shift is negative. This frequency shift is created by the proximity of the target causing a change in the impedance of the antenna. Experiments with a prototype antenna show long-range detection with low power requirements. The detector requires only one loop with one winding which is used for both transmit and receive. This allows for a metal/anomaly detector with a very simple design. The design is lightweight and, depending on loop size, significantly increases detection depth performance. In the full paper, modeling and further experimental results will be presented. Performance results for various types of soil and for different types of targets are presented.

Detection of homemade explosive (HME) containing ammonium nitrate (AN) in the subsurface is of great interest to the US military and its coalition partners. Due to the hygroscopy of AN, this HME is expected to be an intermediate electrical conductivity material (IECM), defined here as one having electrical conductivity greater than soils, which have conductivities 0.1 to 1000 mS•m⁻¹ but less than metals, which have electrical conductivities on the order of 10 MS•m⁻¹. Our preliminary experimental and numerical modeling have established that AN-containing HME in the subsurface can, in all likelihood, be detected by electromagnetic exploration geophysics techniques, specifically by ground penetrating radar (GPR) and by electromagnetic induction (EMI). The electromagnetic induction signatures of HME for these techniques are distinctive. For example, in the case of EMI, the maximum quadrature response frequencies for IECM targets have been found to be greater than 100 kHz.

The ability to detect and dispose of buried mines requires effective means by which to discriminate between hazardous targets and benign clutter. In that regard, wide-band electromagnetic induction (EMI) sensors have shown significant promise in their ability to classify buried metallic objects based on their response to illumination by a time-varying magnetic field. A target's scattered response may be expressed compactly in its magnetic polarizability dyadic, a form that describes the reaction of the scatterer to an arbitrary magnetic field. More specifically, a singularity expansion of the polarizability dyadic has been shown to allow for robust target discrimination. Eddy currents in a 2D surface may be expressed using a scalar stream function, an approach which is powerful since the solenoidality of the current density is enforced trivially. Using this formulation, it is possible to arrive at a modal eigenvalue expansion for the magnetic polarizability that is in the form of the singularity expansion.

Experimental data measured at a field test site with a broadband electromagnetic induction (EMI) sensor are presented. The system is an improved version of the Georgia Tech EMI system developed over the past several years. The system operates over a 300 to one bandwidth and is more sensitive while being more power efficient than earlier systems. Data measured with the system will be presented with an emphasis on features in the data that can be used to separate metallic targets from the soil response and to discriminate between certain classes of metallic targets.

Electromagnetic induction (EMI) systems often employ separate transmit and receive coils that are placed in close proximity to one another. This placement is desirable for packaging reasons, but it makes achieving minimal mutual coupling between the two coils difficult. An iterative-convex optimization procedure that allows coils to be designed for both minimum mutual coupling and maximum on-axis sensitivity is demonstrated, and then this procedure is used to create a pair of double-sided, non-uniformly wound spiral coils, which improve upon previous coils by making the transmit coil resistance uniform. The effects of capacitive loading within the coils are explored, and results showing that the coils perform as expected at low frequency are presented.

Standard protocol for detection and classification of Unexploded Ordnance (UXO) comprises a two-step process that includes an initial digital geophysical mapping (DGM) survey to detect magnetic field anomalies followed by a cued survey at each anomaly location that enables classification of these anomalies. The initial DGM survey is typically performed using a low resolution single axis electromagnetic induction (EMI) sensor while the follow-up cued survey requires revisiting each anomaly location with a multi-axis high resolution EMI sensor. The DGM survey comprises data collection in tightly spaced transects over the entire survey area. Once data collection in this area is complete, a threshold analysis is applied to the resulting magnetic field anomaly map to identify anomalies corresponding to potential targets of interest (TOI). The cued sensor is deployed in static mode where this higher resolution sensor is placed over the location of each anomaly to record a number of soundings that may be stacked and averaged to produce low noise data. These data are of sufficient quality to subsequently classify the object as either TOI or clutter. While this approach has demonstrated success in producing effective classification of UXO, conducting successive surveys is time consuming. Additionally, the low resolution of the initial DGM survey often produces errors in the target picking process that results in poor placement of the cued sensor and often requires several revisits to the anomaly location to ensure adequate characterization of the target space. We present data and test results from an advanced multi-axis EMI sensor optimized to provide both detection and classification from a single survey. We demonstrate how the large volume of data from this sensor may be used to produce effective detection and classification decisions while only requiring one survey of the munitions response area.

This paper extends a previously-introduced method for automatic classification of Unexploded Ordnance (UXO) across several datasets from live sites. We used the MetalMapper sensor, from which extrinsic and intrinsic parameters are determined by the combined Differential Evolution (DE) and Ortho-Normalized Volume Magnetic Source (ONVMS) algorithms. The inversion provides spatial locations and intrinsic time-series total ONVMS principal eigenvalues. These are fit to a power-decay empirical model, providing dimensionality reduction to 3 coefficients (k, b, and g) for polarizability decay. Anomaly target features are grouped using the unsupervised clustering Weighted-Pair Group Method with Averaging (WPGMA) algorithm. Central elements of each cluster are dug, and the results are used to train the next round of dig requests. A Naive Bayes classifier is used as a supervised learning algorithm, in which the product of each feature's independent probability density represents each class of UXO in the feature space. We request ground truths for anomalies in rounds, until there are no more Targets of Interest (TOI) in consecutive requests. This fully automatic procedure requires no expert intervention, saving time and money. Naive Bayes outperformed previous efforts with Gaussian Mixture Models(GMM) in all cases.

This paper presents a combined joint diagonalization (JD) and multiple signal classification (MUSIC) algorithm for estimating subsurface objects locations from electromagnetic induction (EMI) sensor data, without solving ill-posed inverse-scattering problems. JD is a numerical technique that finds the common eigenvectors that diagonalize a set of multistatic response (MSR) matrices measured by a time-domain EMI sensor. Eigenvalues from targets of interest (TOI) can be then distinguished automatically from noise-related eigenvalues. Filtering is also carried out in JD to improve the signal-to-noise ratio (SNR) of the data. The MUSIC algorithm utilizes the orthogonality between the signal and noise subspaces in the MSR matrix, which can be separated with information provided by JD. An array of theoreticallycalculated Green’s functions are then projected onto the noise subspace, and the location of the target is estimated by the minimum of the projection owing to the orthogonality. This combined method is applied to data from the Time-Domain Electromagnetic Multisensor Towed Array Detection System (TEMTADS). Examples of TEMTADS test stand data and field data collected at Spencer Range, Tennessee are analyzed and presented. Results indicate that due to its noniterative mechanism, the method can be executed fast enough to provide real-time estimation of objects’ locations in the field.

The remediation of explosive remnants of war (ERW) and associated unexploded ordnance (UXO) has seen improvements through the injection of modern technological advances and streamlined standard operating procedures. However, reliable and cost-effective detection and geophysical mapping of sites contaminated with UXO such as cluster munitions, abandoned ordnance, and improvised explosive devices rely on the ability to discriminate hazardous items from metallic clutter. In addition to anthropogenic clutter, handheld and vehicle-based metal detector systems are plagued by natural geologic and environmental noise in many post conflict areas. We present new and advanced electromagnetic induction (EMI) technologies including man-portable and towed EMI arrays and associated data processing software. While these systems feature vastly different form factors and transmit-receive configurations, they all exhibit several fundamental traits that enable successful classification of EMI anomalies. Specifically, multidirectional sampling of scattered magnetic fields from targets and corresponding high volume of unique data provide rich information for extracting useful classification features for clutter rejection analysis. The quality of classification features depends largely on the extent to which the data resolve unique physics-based parameters. To date, most of the advanced sensors enable high quality inversion by producing data that are extremely rich in spatial content through multi-angle illumination and multi-point reception.

The appearance of next-generation EMI sensors has been accompanied by the development of advanced EMI models and new interpretation and inversion schemes that take advantage of the richness and diversity of the data provided by these instruments. The technologies have been successfully tested in various scenarios, including ESTCP live-UXO classification studies, and have demonstrated superb classification performances. The studies have shown that the system’s ability to detect and classify small targets (i.e., calibers ranging from 20 to 60 mm) and deep targets (burial depth more than 11 times the target’s diameter) is still a challenging problem when an existing system is used. To overcome this problem, first the standard approach is analyzed, then targets detections are studied for different transmitter coil combinations and transmitter current magnitudes. The results are validated experimentally. The studies are done for a 37mm projectile placed at 42cm and 86 cm under the 2×2 TEMTADS instrument. The target detection and classification performances are illustrated for 6, 11 and 14 Ampere Tx currents using the joint diagonalization and ortho normalized volume magnetic source techniques.

One of the most challenging aspects of survey data processing is target selection. The fundamental input for the classification is dynamic data collected along survey lines. These data are different from the static data obtained in cued mode and used for target classification. Survey data are typically collected using just one transmitter loop (the Z-axis loop) and feature short data point collection times and short decay transience. The collection intervals for each data point are typically 0.1 s, and the signal repetition rates are typically 90 or 270 Hz (in other words, the transient decay times are 2.7 ms or 0.9 ms). Reliable classification requires multiple side/angle illumination; i.e., to conduct reliable classification it is necessary to combine and jointly invert multiple data points. However, picking data points that provide optimal information for classifying targets is a difficult task. The traditional method plots signal amplitudes on a 2D map and picks peaks of signal level without properly accounting for the underlying physics. In this paper, the joint diagonalization is applied to survey data sets to improve data pre-processing and target picking. The JD technique is an EMI data analysis and target classification technique and is applicable for all next-generation multi-static array EMI sensors. The method extracts multi-static response data matrix eigenvalues. The eigenvalues are main characteristics of the data. Recent studies have demonstrated that the method has great potential to quickly estimate the number of potential targets and moreover classify these targets at the data pre-processing stage, in real time and without the need for a forward model. Another advantage of JD is that it provides the ability to separate signal from noise making it possible to de-noise data without distorting the signal due to the targets. In this paper the JD technique is used to process dynamic data collected at South West Proving Ground and Aberdeen Proving Ground (APG) sites using the 2 × 2 TEMTADS and OPTEMA systems, respectively. The joint eigenvalues are extracted as functions of time for each data point and summed/stacked together before being used to create detection maps. Once targets are detected, a set of data is chosen for each anomaly and inverted using the ortho-normalized volume magnetic source technique.

Metal detector coil design has been matured to great extent over the years. However, vehicle mounted or remotely operated metal detectors require different specifications and these specifications dictate multiple transmit and receive coils operating in various settings. Unlike handheld operation, detector is more susceptible to metallic body of the vehicle. Moreover, ground calibration is also different than handheld devices. Coil geometries and intercoupling between them play a significant role in system design and performance. In this study, we study different coil geometries for vehicular applications. Starting from well-known coil geometries, we placed coils on dry, wet and ferrous soil to understand the interaction mechanism. Simulation studies are performed in frequency domain but the results are all applicable to time domain pulse based detector systems.

Buried threat detection system (e.g., GPR, FLIR, EMI) performance can be summarized through two related statistics: the probability of detection (PD), and the false alarm rate (FAR). These statistics impact system rate of forward advance, clearance probability, and the overall usefulness of the system. Understanding system PD and FAR for each target type of interest is fundamental to making informed decisions regarding system procurement and deployment. Since PD and FAR cannot be measured directly, proper experimental design is required to ensure that estimates of PD and FAR are accurate. Given an unlimited number of target emplacements, estimating PD is straightforward. However in realistic scenarios with constrained budgets, limited experimental collection time and space, and limited number of targets, estimating PD becomes significantly more complicated. For example, it may be less expensive to collect data over the same exact target emplacement multiple times than to collect once over multiple unique target emplacements. Clearly there is a difference between the quantity and value of the information obtained from these two experiments (one collection over multiple objects, and multiple collections over one particular object). This work will clarify and quantify the amount of information gained from multiple data collections over one target compared to collecting over multiple unique target burials. Results provide a closed-form solution to estimating the relative value of collecting multiple times over one object, or emplacing a new object, and how to optimize experimental design to achieve stated goals and simultaneously minimize cost.

A broadband arbitrary polarization antenna system for ground penetrating radar applications is modified to improve its performance. The antenna system uses four, crossed, resistive-vee dipole (RVD) antennas operating bistaticly to measure the simultaneous transmission and reception of multiple polarizations. The RVD has low self clutter, low radar cross section and wideband performance. The RVD is a linearly polarized antenna, but other polarizations can be synthesized through the use of two orthogonal RVDs to transmit or receive orthogonal field components. The antenna system is able to distinguish rotationally symmetric and linear targets with its ability to transmit and receive both senses of circular polarization. For example, linear targets such as wires or pipes can be identified by even scattering of both senses of circular polarization. The RVDs in the previous RVD-based CP (circularly polarized) antenna were not designed for CP synthesis. The shape and resistive profile of the RVD were modified to improve dual CP performance. The design of the RVD was optimized through simulation to improve CP synthesis and forward gain, while maintaining low self clutter, low radar cross section, and wide bandwidth. Additional simulations demonstrate that the improvements to the RVD may help to correctly discriminate targets based on their geometries.

Field experience has shown that soil conditions can have large effects on the ability of ground-penetrating radar (GPR) to detect buried targets of interest. The relative permittivity of the soil determines the attenuation of the radar signal. The contrast between the relative permittivity of the soil and the target is critical to determining the strength of the reflection from the target. In this paper, to measure the relative permittivity of the soil and various target fill materials, a microstrip ring resonator is placed in contact with a material medium. The real and imaginary parts of the relative permittivity are determined from (1) changes in resonant frequencies (between 600 MHz and 2 GHz) and (2) the quality factor of the resonator, respectively. Measurement results are compared to data collected by a GPR.

A number of recent algorithms have shown improved performance in detecting buried explosive threats by statistically modeling target responses observed in ground penetrating radar (GPR) signals. These methods extract features from known examples of target responses to train a statistical classifier. The statistical classifiers are then used to identify targets emplaced in previously unseen conditions. Due to the variation in target GPR responses caused by factors such as differing soil conditions, classifiers require training on a large, varied dataset to encompass the signal variation expected in operational conditions. These training collections generally involve burying each target type in a number of soil conditions, at a number of burial depths. The cost associated with both burying the targets, and collecting the data is extremely high. Thus, the conditions and depths sampled cover only a subset of possible scenarios. The goal of this research is to improve the ability of a classifier to generalize to new conditions by deforming target responses in accordance with the physical properties of GPR signals. These signal deformations can simulate a target response under different conditions than those represented in the data collection. This research shows that improved detection performance in previously unseen conditions can be achieved by utilizing deformations, even when the training dataset is limited.

Ground Penetrating Radar (GPR) is a very promising technology for subsurface threat detection. A successful algorithm employing GPR should achieve high detection rates at a low false-alarm rate and do so at operationally relevant speeds. GPRs measure reflections at dielectric boundaries that occur at the interfaces between different materials. These boundaries may occur at any depth, within the sensor's range, and furthermore, the dielectric changes could be such that they induce a 180 degree phase shift in the received signal relative to the emitted GPR pulse. As a result of these time-of-arrival and phase variations, extracting robust features from target responses in GPR is not straightforward. In this work, a method to mitigate polarity and alignment variations based on an expectation-maximization (EM) principal-component analysis (PCA) approach is proposed. This work demonstrates how model-based target alignment can significantly improve detection performance. Performance is measured according to the improvement in the receiver operating characteristic (ROC) curve for classification before and after the data is properly aligned and phase-corrected.

Target detection algorithms for ground penetrating radar (GPR) data typically calculate local statistics for the background data surrounding a test sample as a means to assess changes in the data from background. To ensure that the local statistics are indicative of only the background data and not the data due to a potential target, a guard-band is employed to prohibit the data near the test sample from being used in the calculations. The selection of the guard-band can greatly impact performance, and the value chosen should be based on the expected size of a target response, which is a challenging task when the target population varies greatly. This work develops a robust Bayesian approach to target detection that does not require selection of a guard-band. By modeling the data using Students t distribution rather than a Gaussian distribution, an inference algorithm is developed that automatically identifies outliers from the background and excludes them while calculating the local statistics. The algorithm that was developed is applied to handheld GPR data, where it is shown to provide improved performance over any particular selection of a guard-band.

Down-looking ground-penetrating radar (DLGPR) has been used extensively for buried target detection. For operational implementations, the sensor is used in direct-detection mode, where algorithms process data while the system moves down roadways. Decisions are made before a system passes over the target. Change detection works by passing over an area before and after targets are buried. By comparing before-and-after data, change detection can improve DLGPR performance, but it also has inherent operational limitations. Performance enhancements include mitigating the effects of anomalies not associated with targets and increasing the detection probabilities of deeper targets through indirect means. In the latter case, deeply buried targets that do not appear in the GPR data can be indirectly detected using change detection methods if the patch of ground where the target is buried has been significantly modified from its original undisturbed state. In this paper, we explore decision-based change-detection approaches for enhancing the performance of a DLGPR system and enumerate the limitations of the approach.

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%.

Ground Penetrating radar (GPR) is a commonly used modality for the detection of buried threats. This work explores two approaches for buried threat detection in GPR data that we refer to as the hyperbolic filter and PLSDA filter algorithms. The hyperbolic filter algorithm leverages the hyperbolic shape of buried threat GPR responses, while the PLSDA algorithm uses a PLSDA linear classifier to learn a filter based on classifier weights. A hyperbolic filter is trained and optimized by doing a grid search over a set of hyperbola parameters. The PLSDA filter is generated by aligning GPR data and training PLSDA weights on that feature space. The correlation between each filter and the 2D GPR data provides information regarding the presence of buried threats. The PLSDA and hyperbolic filters were generated for a data set containing multiple target types. Both PLSDA and hyperbolic filters outperformed a prescreener for target subsets, and performed similarly over all target types. Relative to one another, both PLSDA and hyperbolic filters performed equally well. PLSDA filters, however, can be trained much faster than the corresponding exhaustive search needed by the hyperbolic filter.

We present a fusion method, based on fuzzy inference, for detecting buried objects using ground-penetrating radar (GPR) data. The GPR sensor generates 3-dimensional data that correspond to depth, down-track, and cross-track. Most discrimination algorithms process only 2-D slices of the 3-D cube: (down-track, depth) or (cross-track, depth). The performance of the down-track and cross-track discrimination algorithms can vary significantly depending on the target shape, burial orientation, and other environmental conditions. In some cases, these algorithms can provide complementary evidence, while in other cases they provide contradicting evidence. Thus, effective fusion of these algorithms can achieve higher probability of detection with fewer false alarms. The proposed fusion method is capable of learning meaningful and simple fuzzy rules for different regions of the input space, generated by partial confidence values of the different discriminators as well as additional background information. To learn the rules, first, the input space is partitioned to identify local contexts. Second, input membership functions are learned based on the distribution of the partial confidence values of the individual discriminators within each context. Third, output membership functions are generated by considering the relative numbers of targets and non-targets within each context. Finally, the input and output membership functions are combined into a Mamdani-type fuzzy inference system. The output of the learning process is a fuzzy rule base adapted to different contexts. Results on large and diverse GPR data collections show that the proposed fusion can identify local, simple, and meaningful rules capable of non-linear fusion of different discriminators. We also show that the proposed fuzzy inference outperforms commonly used fusion methods.

Ground penetrating radar (GPR) is a popular sensing modality for buried threat detection that offers low false alarm rates (FARs), but suffers from a short detection stopping or standoff distance. This short stopping distance leaves little time for the system operator to react when a threat is detected, limiting the speed of advance. This problem arises, in part, because of the way GPR data is typically processed. GPR data is first prescreened to reduce the volume of data considered for higher level feature-processing. Although fast, prescreening introduces latency that delays the feature processing and lowers the stopping distance of the system. In this work we propose a novel sensor fusion framework where a forward looking infrared (FLIR) camera is used as a prescreener, providing suspicious locations to the GPRbased system with zero latency. The FLIR camera is another detection modality that typically yields a higher FAR than GPR while offering much larger stopping distances. This makes it well-suited in the role of a zero-latency prescreener. In this framework, GPR-based feature processing can begin without any latency, improving stopping distances. This framework was evaluated using well-known FLIR and GPR detection algorithms on a large dataset collected at a Western US test site. Experiments were conducted to investigate the tradeoff between early stopping distance and FAR. The results indicate that earlier stopping distances are achievable while maintaining effective FARs. However, because an earlier stopping distance yields less data for feature extraction, there is a general tradeoff between detection performance and stopping distance.

In this paper, microwave holography is considered as a tool to obtain high resolution images of shallowly buried objects. Signal acquisition is performed at multiple frequencies on a grid using a two-dimensional mechanical scanner moving a single transceiver over an area of interest in close proximity to the surface. The described FFT-based reconstruction technique is used to obtain a stack of plan view images each using only one selected frequency from the operating waveband of the radar. The extent of a synthetically-formed aperture and the signal wavelength define the plan view resolution, which at sounding frequencies near 7 GHz amounts to 2 cm. The system has a short depth of focus which allows easy selection of proper focusing plane. The small distance from the buried objects to the antenna does not prevent recording of clean images due to multiple reflections (as happens with impulse radars). The description of the system hardware and signal processing technique is illustrated using experiments conducted in dry sand. The microwave images of inert anti-personnel mines are demonstrated as examples. The images allow target discrimination based on the same visually-discernible small features that a human observer would employ. The demonstrated technology shows promise for modification to meet the specific practical needs required for humanitarian demining or in multi-sensor survey systems.

Buried explosive hazards (BEHs) have been, and continue to be, one of the most deadly threats in modern conflicts. Current handheld sensors rely on a highly trained operator for them to be effective in detecting BEHs. New algorithms are needed to reduce the burden on the operator and improve the performance of handheld BEH detectors. Traditional anomaly detection and discrimination algorithms use “hand-engineered” feature extraction techniques to characterize and classify threats. In this work we use a Deep Belief Network (DBN) to transcend the traditional approaches of BEH detection (e.g., principal component analysis and real-time novelty detection techniques). DBNs are pretrained using an unsupervised learning algorithm to generate compressed representations of unlabeled input data and form feature detectors. They are then fine-tuned using a supervised learning algorithm to form a predictive model. Using ground penetrating radar (GPR) data collected by a robotic cart swinging a handheld detector, our research demonstrates that relatively small DBNs can learn to model GPR background signals and detect BEHs with an acceptable false alarm rate (FAR). In this work, our DBNs achieved 91% probability of detection (P_d) with 1.4 false alarms per square meter when evaluated on anti-tank and anti-personnel targets at temperate and arid test sites. This research demonstrates that DBNs are a viable approach to detect and classify BEHs.

The U.S. Department of Defense Humanitarian Demining Research and Development (HD R&D) Program, Non-Intrusive Inspection Technology (NIITEK), Inc. and The HALO Trust have over the last decade funded, developed and tested various prototype vehicle mounted ground penetrating radar (GPR) systems named the Mine Stalker. The HD R&D Program and NIITEK developed the Mine Stalker to detect low metal anti-tank (LM-AT) mines in roads. The country of Angola is severely affected by LM-AT mines in and off road, some of which are buried beyond the effective range of detection sensors current used in country. The threat from LM-AT mines such as the South African Number 8 (No. 8) and the Chinese Type 72 (72AT) still persist from Angola’s 30 years of civil war. These LM-AT threats are undetectable at depths greater than 5 to 10 centimeters using metal detection technology. Clearing commerce routes are a critical requirement before Angola can rebuild its infrastructure and improve safety conditions for the local populace. The Halo Trust, a non-governmental demining organization (NGO) focused on demining and clearance of unexploded ordnance (UXO), has partnered with the HD R&D Program to conduct an operational field evaluation (OFE) of the Mine Stalker III (MS3) in Angola. Preliminary testing and training efforts yielded encouraging results. This paper presents a review of the data collected, testing results, system limitations and deficiencies while operating in a real world environment. Our goal is to demonstrate and validate this technology in live minefield environments, and to collect data to prompt future developments to the system.

Standoff Raman imaging systems have shown the ability to detect single explosives particles. However, in many cases, the laser intensities needed restrict the applications where they can be safely used. A new generation imaging Raman system has been developed based on a 355 nm UV laser that, in addition to eye safety, allows discrete and invisible measurements. Non-dangerous exposure levels for the eye are several orders of magnitude higher in UVA than in the visible range that previously has been used. The UV Raman system has been built based on an UV Fabry-Perot Interferometer (UV-FPI) developed by VTT. The design allows for precise selection of Raman shifts in combination with high out-of-band blocking. The stable operation of the UV-FPI module under varying environmental conditions is arranged by controlling the temperature of the module and using a closed loop control of the FPI air gap based on capacitive measurement. The system presented consists of a 3rd harmonics Nd:YAG laser with 1.5 W average output at 1000 Hz, a 200 mm Schmidt-Cassegrain telescope, UV-FPI filter and an ICCD camera for signal gating and detection. The design principal leads to a Raman spectrum in each image pixel. The system is designed for field use and easy manoeuvring. Preliminary results show that in measurements of <60 s on 10 m distance, single AN particles of <300 μm diameter can be identified.

A new method of explosives detection based on the field asymmetric ion mobility spectrometry (FAIMS) and ionization by an excimer emitter has been developed jointly with a portable detector. The excimer emitter differs from usual UVionizing lamps by mechanism of emitting, energy and spectral characteristics. The developed and applied Ar₂-excimer emitter has the working volume of 1 cm³, consuming power 0.6 W, the energy of photons of about 10 eV (λ=126 nm), the FWHM radiation spectrum of 10 nm and emits more than 10¹⁶ photon per second that is two orders of magnitude higher than UV-lamp of the same working volume emits. This also exceeds by an order of magnitude the quantity of photons per second for 10-Hz solid state YAG:Nd³⁺ - laser of 1mJ pulse energy at λ=266 nm that is also used to ionize the analyte. The Ar₂-excimer ionizes explosives by direct ionization mechanism and through ionization of organic impurities. The developed Ar₂-excimer-based ion source does not require cooling due to low level discharge current of emitter and is able to work with no repair more than 10000 hrs. The developed excimer-based explosives detector can analyze both vapors and traces of explosives. The FAIMS spectra of the basic types of explosives like trinitrotoluene (TNT), cyclotrimethylenetrinitramine (RDX), dinitrotoluene (DNT), cyclotetramethylenetetranitramine (HMX), nitroglycerine (NG), pentaerythritol tetranitrate (PETN) under Ar₂-excimer ionization are presented. The detection limit determined for TNT vapors equals 1x10^-14 g/cm³, for TNT traces- 100 pg.

The development of new techniques for the detection of homemade explosive devices is an area of intense research for the defense community. Such sensors must exhibit high selectivity to detect explosives and/or explosives related materials in a complex environment. Spectroscopic techniques such as FTIR are capable of discriminating between the volatile components of explosives; however, there is a need for less expensive systems for wide-range use in the field. To tackle this challenge we are investigating the use of multiple, overlapping, broad-band infrared (IR) filters to enable discrimination of volatile chemicals associated with an explosive device from potential background interferants with similar chemical signatures. We present an optical approach for the detection of fuel oil (the volatile component in ammonium nitrate-fuel oil explosives) that relies on IR absorption spectroscopy in a laboratory environment. Our proposed system utilizes a three filter set to separate the IR signals from fuel oil and various background interferants in the sample headspace. Filter responses for the chemical spectra are calculated using a Gaussian filter set. We demonstrate that using a specifically chosen filter set enables discrimination of pure fuel oil, hexanes, and acetone, as well as various mixtures of these components. We examine the effects of varying carrier gasses and humidity on the collected spectra and corresponding filter response. We study the filter response on these mixtures over time as well as present a variety of methods for observing the filter response functions to determine the response of this approach to detecting fuel oil in various environments.

The proliferation of Radio Controlled Improvised Explosive Devices (RC-IED) is a growing threat around the world. The ease of construction and low cost of these devices are transforming common things in lethal tramps. The fight against this threats normally involves the use of sophisticated and expensive equipment of Electronic Warfare based on high speed DSP systems, just to detect the presence of detonation signals. In this work is showed how to find activation signals based on the characteristic of the power in a specific band and the previous knowledge about the detonation signals. As proof of concept we have taken the information about the RC-IEDs used in the Colombian conflict and develop an algorithm to find detonation signals based on the measured power in frequencies between 136 MHz and 174 MHz (2 meter civil band)

An improvised explosive device (IED) is a bomb constructed and deployed in a non-standard manor. Improvised means that the bomb maker took whatever he could get his hands on, making it very hard to predict and detect. Nevertheless, the matters in which the IED’s are deployed and used, for example as roadside bombs, follow certain patterns. One possible approach for early warning is to record the surroundings when it is safe and use this as reference data for change detection. In this paper a LADAR-based system for IED detection is presented. The idea is to measure the area in front of the vehicle when driving and comparing this to the previously recorded reference data. By detecting new, missing or changed objects the system can make the driver aware of probable threats.

It is well-known that a pattern recognition system is only as good as the features it is built upon. In the fields of image processing and computer vision, we have numerous spatial domain and spatial-frequency domain features to extract characteristics of imagery according to its color, shape and texture. However, these approaches extract information across a local neighborhood, or region of interest, which for target detection contains both object(s) of interest and background (surrounding context). A goal of this research is to filter out as much task irrelevant information as possible, e.g., tire tracks, surface texture, etc., to allow a system to place more emphasis on image features in spatial regions that likely belong to the object(s) of interest. Herein, we outline a procedure coined soft feature extraction to refine the focus of spatial domain features. This idea is demonstrated in the context of an explosive hazards detection system using forward looking infrared imagery. We also investigate different ways to spatially contextualize and calculate mathematical features from shearlet filtered candidate image chips. Furthermore, we investigate localization strategies in relation to different ways of grouping image features to reduce the false alarm rate. Performance is explored in the context of receiver operating characteristic curves on data from a U.S. Army test site that contains multiple target and clutter types, burial depths, and times of day.

In this paper we present an approach that uses Genetic Programming (GP) to evolve novel feature extraction algorithms for greyscale images. Our motivation is to create an automated method of building new feature extraction algorithms for images that are competitive with commonly used human-engineered features, such as Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG). The evolved feature extraction algorithms are functions defined over the image space, and each produces a real-valued feature vector of variable length. Each evolved feature extractor breaks up the given image into a set of cells centered on every pixel, performs evolved operations on each cell, and then combines the results of those operations for every cell using an evolved operator. Using this method, the algorithm is flexible enough to reproduce both LBP and HOG features. The dataset we use to train and test our approach consists of a large number of pre-segmented image “chips” taken from a Forward Looking Infrared Imagery (FLIR) camera mounted on the hood of a moving vehicle. The goal is to classify each image chip as either containing or not containing a buried object. To this end, we define the fitness of a candidate solution as the cross-fold validation accuracy of the features generated by said candidate solution when used in conjunction with a Support Vector Machine (SVM) classifier. In order to validate our approach, we compare the classification accuracy of an SVM trained using our evolved features with the accuracy of an SVM trained using mainstream feature extraction algorithms, including LBP and HOG.

A convolutional neural network (CNN) approach to recognition of buried explosive hazards in forward-looking long-wave infrared (FL-LWIR) imagery is presented. The convolutional filters in the first layer of the network are learned in the frequency domain, making enforcement of zero-phase and zero-dc response characteristics much easier. The spatial domain representations of the filters are forced to have unit l2 norm, and penalty terms are added to the online gradient descent update to encourage orthonormality among the convolutional filters, as well smooth first and second order derivatives in the spatial domain. The impact of these modifications on the generalization performance of the CNN model is investigated. The CNN approach is compared to a second recognition algorithm utilizing shearlet and log-gabor decomposition of the image coupled with cell-structured feature extraction and support vector machine classification. Results are presented for multiple FL-LWIR data sets recently collected from US Army test sites. These data sets include vehicle position information allowing accurate transformation between image and world coordinates and realistic evaluation of detection and false alarm rates.

In this paper, we propose a reinforcement random forest algorithm as a novel approach to detect unpaved road regions at stand-off distances. A random forest classifier is used to differentiate between road and non-road pixels/patches without over fitting the training data. Utilizing a reinforcement technique, the algorithm can handle foreign objects that we encounter in real world driving. Furthermore, classifying road patches at different distances generates multiple levels of road agreement for each pixel within the image. Using different threshold values of this agreement level provides adaptability to the road finding results. The selection of low threshold values produces better detection rates but also increases false alarms. On the other hand, high threshold values lower the detection rate and decreases false detections. In our experiments, the proposed algorithm is tested on color video of unpaved road in an arid environment.

An approach to image classification based on the analysis of the network of points generated by an image feature detection algorithm has been proposed. This network-based approach looks at the networks produced by two images and scale and then compare them, making a classification decision. This paper considers techniques to handle the problem posed by input images that are obscured or in which the target is partially covered. These approaches are compared with the base algorithm to assess the impact on performance in the general case, obscured scenarios and obstructed scenarios.

Imaging flash LIDAR (LIght Detection and Ranging) is an effective method for airborne searches of the ocean surface and subsurface volume. The performance of ocean LIDAR depends strongly on the sea surface (e.g., waves, whitecaps, and flotsam), water turbidity, and the characteristics of the objects of interest. Cost-effective design of the LIDAR system and processing algorithms requires a modeling capability that can deal with the physics of light propagation through the air-water interface, into the ocean, and back to the LIDAR receiver. 3DLASE-M is a physics-based LIDAR simulator that yields high-fidelity images for three-dimensional algorithm development and performance predictions.

A polarization thermal sensor is presented that has ability to detect manmade objects even when there is little or no thermal contrast, or when the object is obscured by clutter or vegetation. Imagery will be shown for scenarios in which polarization contrast remains high during periods of zero thermal contrast. Also, data will be presented that shows the detection of manmade targets in natural clutter, long after thermal equilibrium of the target with the background has been established.

Undisturbed soil generally consists of larger silica particles where the smaller particles have been removed from the surface by weathering, i.e., rain. Large silica particles exhibit a significant reflectance in the 7 – 12 μm spectral region called the Reststrahlen band. It has been demonstrated that the Reststrahlen band intensity is proportional to the particle size distribution of silica in the soil; as the proportion of small to large silica particles increases the Reststrahlen band also decreases in intensity. When soil is disturbed, for example when an object is buried, the distribution of silica particles is changed such that the “new” surface consists of a greater proportion of small to large silica particles. The increased number of small silica particles decreases the intensity of the Reststrahlen band therefore enabling detection of buried objects by imaging in the 7 – 12 μm spectral range. There is a second reflectance band in the 17 – 25 μm spectral region which has not been used for detection of buried objects. We investigate the behavior of this second Reststrahlen band for disturbed earth and evaluate its use for enhanced detection of buried objects when combined with the first Reststrahlen band.