In this work we present a new concept we call self-assessment which leverages techniques from the self-supervision literature to estimate algorithm performance and certainty in real-time. Our self-assessment approach enables AI/ML systems to determine what they know, how well they know it, identify what is fundamentally knowable about a scene, and highlight confusing or un-reliable data for further investigation. This approach has applications to identifying out-of-domain data or flagging unexpected changes in operational contexts that can otherwise reduce AI/ML trustworthiness.
Improved performance in the discrimination of buried threats using Ground Penetrating Radar (GPR) data has recently been achieved using features developed for applications in computer vision. These features, designed to characterize local shape information in images, have been utilized to recognize patches that contain a target signature in two-dimensional slices of GPR data. While these adapted features perform very well in this GPR application, they were not designed to specifically differentiate between target responses and background GPR data. One option for developing a feature specifically designed for target differentiation is to manually design a feature extractor based on the physics of GPR image formation. However, as seen in the historical progression of computer vision features, this is not a trivial task. Instead, this research evaluates the use of convolutional neural networks (CNNs) applied to two-dimensional GPR data. The benefit of using a CNN is that features extracted from the data are a learned parameter of the system. This has allowed CNN implementations to achieve state of the art performance across a variety of data types, including visual images, without the need for expert designed features. However, the implementation of a CNN must be done carefully for each application as network parameters can cause performance to vary widely. This paper presents results from using CNNs for object detection in GPR data and discusses proper parameter settings and other considerations.
A goal of ground penetrating radar (GPR) preprocessing is to distinguish background from data containing explosive threats. This is commonly achieved by performing depth-dependent mean and standard deviation normalization, where the mean and standard deviation are computed on background data. Under the assumption that data with explosive threats have different statistical characteristics than the background/clutter, after normalization explosive threat data will have larger absolute normalized scores than the background/clutter. An underlying problem is determining which data to compute the background mean and standard deviation statistics over. Often the background statistics are computed over a moving window, which is centered at the location of interest and has a predetermined guard band, a region of data that is ignored. However, buried explosive threats vary considerably in their shapes and more importantly sizes subsequently, the size of the GPR responses from these objects are considerably varied. We examine a number of additional detection methods that utilize Robust Principal Component Analysis (RPCA), where RPCA decomposes the data into low-rank and sparse components. Intuitively, the low-rank component should capture the background data and the sparse should capture the anomalous explosive threat response. We find that detection performance using energy- and shape-based detection algorithms improves when using RPCA preprocessing.
KEYWORDS: General packet radio service, Ground penetrating radar, Signal processing, Target detection, Classification systems, Data modeling, Receivers, Image processing, Image classification, Binary data
Ground-penetrating radar (GPR) technology has proven capable of detecting buried threats. The system relies on a binary classifier that is trained to distinguish between two classes: a target class, encompassing many types of buried threats and their components; and a nontarget class, which includes false alarms from the system prescreener. Typically, the training process involves a simple partition of the data into these two classes, which allows for straightforward application of standard classifiers. However, since training data is generally collected in fully controlled environments, it includes auxiliary information about each example, such as the specific type of threat, its purpose, its components, and its depth. Examples from the same specific or general type may be expected to exhibit similarities in their GPR data, whereas examples from different types may differ greatly. This research aims to leverage this additional information to improve overall classification performance by fusing classifier concepts for multiple groups, and to investigate whether structure in this information can be further utilized for transfer learning, such that the amount of expensive training data necessary to learn a new, previously-unseen target type may be reduced. Methods for accomplishing these goals are presented with results from a dataset containing a variety of target types.
KEYWORDS: Land mines, General packet radio service, Data modeling, Detection and tracking algorithms, Target detection, Data analysis, Process modeling, Ground penetrating radar, Remote sensing, Sensors
Buried threat detection algorithms in Ground Penetrating Radar (GPR) measurements often utilize a statistical classifier to model target responses. There are many different target types with distinct responses and all are buried in a wide range of conditions that distort the target signature. Robust performance of this classifier requires it to learn the distinct responses of target types while accounting for the variability due to the physics of the emplacement. In this work, a method to reduce certain sources of excess variation is presented that enables a linear classifier to learn distinct templates for each target type’s response despite the operational variability. The different target subpopulations are represented by a Gaussian Mixture Model (GMM). Training the GMM requires jointly extracting the patches around target responses as well as learning the statistical parameters as neither are known a priori. The GMM parameters and the choice of patches are determined by variational Bayesian methods. The proposed method allows for patches to be extracted from a larger data-block that only contain the target response. The patches extracted from this method improve the ROC for distinguishing targets from background clutter compared to the patches extracted using other patch extraction methods aiming to reduce the operational variability.
Forward-looking ground penetrating radar (FLGPR) is a remote sensing modality that has recently been investigated for buried threat detection. FLGPR offers greater standoff than other downward-looking modalities such as electromagnetic induction and downward-looking GPR, but it suffers from high false alarm rates due to surface and ground clutter. A stepped frequency FLGPR system consists of multiple radars with varying polarizations and bands, each of which interacts differently with subsurface materials and therefore might potentially be able to discriminate clutter from true buried targets. However, it is unclear which combinations of bands and polarizations would be most useful for discrimination or how to fuse them. This work applies sparse structured basis pursuit, a supervised statistical model which searches for sets of bands that are collectively effective for discriminating clutter from targets. The algorithm works by trying to minimize the number of selected items in a dictionary of signals; in this case the separate bands and polarizations make up the dictionary elements. A structured basis pursuit algorithm is employed to gather groups of modes together in collections to eliminate whole polarizations or sensors. The approach is applied to a large collection of FLGPR data for data around emplaced target and non-target clutter. The results show that a sparse structure basis pursuits outperforms a conventional CFAR anomaly detector while also pruning out unnecessary bands of the FLGPR sensor.
In this work, we explore the efficacy of two buried threat detectors on handheld data. The first algorithm is an energy-based algorithm, which computes how anomalous a given A-scan measurement after it is normalized according to its local statistics. It is based on a commonly used prescreener for the Husky Mounted Detection System (HMDS). In the HMDS setting measurements are sampled on a crosstrack-downtrack grid, and sequential measurements are at neighboring downtrack locations. In contrast, in the handheld setting sequential scans are often taken at neighboring crosstrack locations, and neighboring downtrack locations can be hundreds of scans away. In order to include both downtrack and crosstrack information, we compute local statistics over a much larger area than in the HMDS setting. The second algorithm is a shape-based algorithm. Shape Invariant Feature Transform (SIFT) features, which capture the gradient distributions of local patches, are extracted and used to train a non-linear Support Vector Machine (SVM). We found that in terms of AUC, the SIFT-SVM algorithm results in a 2.2% absolute improvement over the energy-based algorithm, with the greatest gains seen at lower false alarm rates.
KEYWORDS: Forward looking infrared, General packet radio service, Sensors, Systems modeling, Data modeling, Computing systems, Detection and tracking algorithms, Target detection, Cameras, Infrared sensors
Many remote sensing modalities have been developed for buried target detection, each one offering its own relative advantages over the others. As a result there has been interest in combining several modalities into a single detection platform that benefits from the advantages of each constituent sensor, without suffering from their weaknesses. Traditionally this involves collecting data continuously on all sensors and then performing data, feature, or decision level fusion. While this is effective for lowering false alarm rates, this strategy neglects the potential benefits of a more general system-level fusion architecture. Such an architecture can involve dynamically changing which modalities are in operation. For example, a large standoff modality such as a forward-looking infrared (FLIR) camera can be employed until an alarm is encountered, at which point a high performance (but short standoff) sensor, such as ground penetrating radar (GPR), is employed. Because the system is dynamically changing its rate of advance and sensors, it becomes difficult to evaluate the expected false alarm rate and advance rate. In this work, a probabilistic model is proposed that can be used to estimate these quantities based on a provided operating policy. In this model the system consists of a set of states (e.g., sensors employed) and conditions encountered (e.g., alarm locations). The predictive accuracy of the model is evaluated using a collection of collocated FLIR and GPR data and the results indicate that the model is effective at predicting the desired system metrics.
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.
KEYWORDS: General packet radio service, Forward looking infrared, Sensors, Standoff detection, Feature extraction, Detection and tracking algorithms, Data processing, Cameras, Ground penetrating radar, Land mines
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.
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.
KEYWORDS: Land mines, General packet radio service, Principal component analysis, Data modeling, Target detection, Sensors, Expectation maximization algorithms, Detection and tracking algorithms, Distortion, Dielectrics
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.
KEYWORDS: Target detection, Data modeling, Sensors, Detection and tracking algorithms, Remote sensing, General packet radio service, Systems modeling, Electromagnetic coupling, Sensing systems, Visualization
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.
KEYWORDS: General packet radio service, Land mines, Dielectrics, Sensors, Data modeling, Ground penetrating radar, Antennas, Signal detection, Explosives, Target detection
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.
Forward Looking Infrared (FLIR) cameras have recently been studied as a sensing modality for use in buried threat detection systems. FLIR-based detection systems benefit from larger standoff distances and faster rates of advance than other sensing modalities, but they also present significant signal processing challenges. FLIR imagery typically yields multiple looks at each surface area, each of which is obtained from a different relative camera pose and position. This multi-look imagery can be exploited for improved performance, however open questions remain as to the best ways to process and fuse such data. Further, the utility of each look in the multi-look imagery is also unclear: How many looks are needed, from what poses, etc? In this work we propose a general framework for processing FLIR imagery wherein FLIR imagery is partitioned according to the particular relative camera pose from which it was collected. Each partition is then projected into a common spatial coordinate system resulting in several distinct images of the surface area. Buried threat detection algorithms can then be applied to each of these resulting images independently, or in aggregate. The proposed framework is evaluated using several detection algorithms on an FLIR dataset collected at a Western US test site and the results indicate that the framework offers significant improvement over detection in the original FLIR imagery. Further experiments using this framework suggest that multiple looks by the FLIR camera can be used to improve detection performance.
KEYWORDS: Land mines, General packet radio service, Fusion energy, Detection and tracking algorithms, Image processing, Sensors, Visualization, Data modeling, Feature extraction, Data processing
Utilizing methods from the image processing and computer vision fields has led to advances in high resolution Ground Penetrating Radar (GPR) based threat detection. By analyzing 2-D slices of GPR data and applying various image processing algorithms, it is possible to discriminate between threat and non-threat objects. In initial attempts to utilize such approaches, object instance-matching algorithms were applied to GPR images, but only limited success was obtained when utilizing feature point methods to identify patches of data that displayed landmine-like characteristics. While the approach worked well under some conditions, the instance-matching method of classification was not designed to identify a type of class, only reproductions of a specific instance. In contrast, our current approach is focused on identifying methods that can account for within-class variations that result from changing target types and varying operating conditions that a GPR system regularly encounters. Image category recognition is an area of research that attempts to account for within class variation of objects within visual images. Instead of finding a reproduction of a particular known object within an image, algorithms for image categorization are designed to learn the qualities of images that contain an instance belonging to a known class. The results illustrate how image category recognition algorithms can be successfully applied to threat identification in GPR data.
KEYWORDS: Land mines, Data modeling, General packet radio service, Statistical modeling, Target detection, Feature extraction, Matrices, Image segmentation, Detection and tracking algorithms, Signal detection
Ground Penetrating Radar (GPR) is a widely used technology for the detection of subsurface buried threats. Although GPR data contains a representation of 3D space, during training, target and false alarm locations are usually only provided in 2D space along the surface of the earth. To overcome uncertainty in target depth location, many algorithms simply extract features from multiple depth regions that are then independently used to make mine/non-mine decisions. A similar technique is employed in hidden Markov models (HMM) based landmine detection. In this approach, sequences of downtrack GPR responses over multiple depth regions are utilized to train an HMM, which learns the probability of a particular sequence of GPR responses being generated by a buried target. However, the uncertainty in object depth complicates learning for discriminating targets/non-targets since features at the (unknown) target depth can be significantly different from features at other depths but in the same volume. To mitigate the negative impact of the uncertainty in object depth, mixture models based on Multiple Instance Learning (MIL) have previously been developed. MIL is also applicable in the landmine detection problem using HMMs because features that are extracted independently from sequences of GPR signals over several depth bins can be viewed as a set of unlabeled time series, where the entire set either corresponds to a buried threat or a false alarm. In this work, a novel framework termed as multiple instance hidden Markov model (MIHMM) is developed. We show that the performance of the proposed approach for discriminating targets from non-targets in GPR data is promising.
Frequency-domain electromagnetic induction (EMI) sensors have been shown to provide target signatures which enable discrimination of landmines from harmless clutter. In particular, frequency-domain EMI sensors are well-suited for target characterization by inverting a physics-based signal model. In many model-based signal processing paradigms, the target signatures can be decomposed into a weighted sum of parameterized basis functions, where the basis functions are intrinsic to the target under consideration and the associated weights are a function of the target sensor orientation. When sensor array data is available, the spatial diversity of the measured signals may provide more information for estimating the basis function parameters. After model inversion, the basis function parameters can form the foundation of model-based classification of the target as landmine or clutter. In this work, sparse model inversion of spatial frequency-domain EMI sensor array data followed by target classification using a statistical model is investigated. Results for data measured with a prototype frequency-domain EMI sensor at a standardized test site are presented. Preliminary results indicate that extracting physics-based features from spatial frequency-domain EMI sensor array data followed by statistical classification provides an effective approach for classifying targets as landmine or clutter.
A method is presented for determining the position and orientation of a vehicle from a single, color video taken
from the hood of the vehicle, for the purpose of assisting its autonomous operation at very high speeds on
rural roads. An implicit perspective transformation allows estimation of the vehicle's orientation and cross-road
image features. From these, an adaptive road model is built and the horizontal position of the vehicle can be
estimated. This method makes very few assumptions about the structure of the road or the path of the vehicle.
In a realistic, simulated environment, good road model construction and vehicle position estimation are achieved
at frame rates suitable for real-time high speed driving.
Frequency-domain electromagnetic induction (EMI) sensors have been shown to provide target signatures which enable
discrimination of landmines from harmless clutter. In particular, frequency-domain EMI sensors are well-suited for target
characterization by inverting a physics-based signal model. In many model-based signal processing paradigms, the target
signatures can be decomposed into a weighted sum of parameterized basis functions, where the basis functions are intrinsic
to the target under consideration and the associated weights are a function of the target sensor orientation. When spatial
data is available, the diversity of the measured signals may provide more information for estimating the basis function
parameters. After model inversion, the basis function parameters can be used as features for classifying the target as
landmine or clutter. In this work, feature extraction from spatial frequency-domain EMI sensor data is investigated. Results
for data measured with a prototype frequency-domain EMI sensor at a standardized test site are presented. Preliminary
results indicate that Structured relevance vector machine (sRVM) regression model inversion using spatial data provides
stable, and sparse, sets of target features.
KEYWORDS: Sensors, Land mines, General packet radio service, Target detection, Statistical analysis, Ground penetrating radar, Dielectrics, Electromagnetism, Detection and tracking algorithms, Head
Ground penetrating radar (GPR) is a commonly employed sensing modality for landmine detection. It has been successfully
deployed in vehicular systems, and is also being integrated into handheld systems. Handheld mine detection systems
are typically deployed in situations where either the terrain or mission renders a vehicular-based system less effective.
Handheld systems are often more compact and maneuverable, but quality of the sensor data may also be more dependent
on the operators experience with and technique in using the system. In particular, the sensor height with respect to the
air-ground interface may be more variable than with a vehicular-based system. This variation in sensor height above the
air-ground interface may have the potential to adversely affect mine detection performance with the GPR sensing modality.
In this work, the effects of operator technique on handheld sensor data quality is investigated, and ground alignment is
explored as a potential approach to reducing variability in the sensor data quality due to operator technique. Results for
data measured with a standard GPR/EMI handheld sensor at a standardized test site are presented.
KEYWORDS: General packet radio service, Land mines, Detection and tracking algorithms, Data fusion, Feature extraction, Metals, Performance modeling, Ground penetrating radar, Sensors, Soil science
Ground-penetrating radar (GPR) is a very useful technology for buried threat detection applications which is
capable of identifying both metallic and non-metallic objects with moderate false alarm rates. Several pattern
classication algorithms have been proposed and evaluated which enable GPR systems to achieve robust per-
formance. However, comparisons of these algorithms have shown that their relative performance varies with
respect to the environmental context under which the GPR is operating. Context-dependent fusion has been
proposed as a technique for algorithm fusion and has been shown to improve performance by exploiting the
dierences in algorithm performance under dierent environmental and operating conditions. Early approaches
to context-dependent fusion clustered observations in the joint condence space of all algorithms and applied
fusion rules within each cluster (i.e., discriminative learning). Later approaches exploited physics-based fea-
tures extracted from the background data to leverage more environmental information, but decoupled context
learning from algorithm fusion (i.e., generative learning). In this work, a Bayesian inference technique which
combines the generative and discriminative approaches is proposed for physics-based context-dependent fusion
of detection algorithms for GPR. The method uses a Dirichlet process (DP) mixture as a model for context, and
relevance vector machines (RVMs) as models for algorithm fusion. Variational Bayes is used as an approximate
inference technique for joint learning of the context and fusion models. Experimental results compare the pro-
posed Bayesian discriminative technique to generative techniques developed in past work by investigating the
similarities and dierences in the contexts learned as well as overall detection performance.
KEYWORDS: Data modeling, Land mines, Process modeling, General packet radio service, Feature extraction, Algorithm development, Machine learning, Computer simulations, Detection and tracking algorithms, Ground penetrating radar
Ground Penetrating Radar (GPR) has been extensively employed as a technology for the detection of subsurface
buried threats. Although vehicular mounted GPRs generate data in three dimensions, alarm declarations are
usually only available in the form of 2-D spatial coordinates. The uncertainty in the depth of the target in the
three dimensional volume of data, and the difficulties associated with automatically localizing objects in depth,
can adversely impact feature extraction and training in some detection algorithms. In order to mitigate the
negative impact of uncertainty in target depth, several algorithms have been developed that extract features from
multiple depth regions and utilize these feature vectors in classification algorithms to perform final mine/nonmine
decisions. However, the uncertainty in object depth significantly complicates learning since features at
the correct target depth are often significantly different from features at other depths but in the same volume.
Multiple Instance Learning (MIL) is a type of supervised learning approach in which labels are available for a
collection of feature vectors but not for individual samples, or in this application, depths. The goal of MIL is
to classify new collections of vectors as they become available. This set-based learning method is applicable in
the landmine detection problem because features that are extracted independently from several depth bins can
be viewed as a set of unlabeled feature vectors, where the entire set either corresponds to a buried threat or
a false alarm. In this work, a novel generative Dirichlet Process Gaussian mixture model for MIL is developed
that automatically infers the number of mixture components required to model the underlying distributions of
mine/non-mine signatures and performs classification using a likelihood ratio test. In this work, we show that the
performance of the proposed approach for discriminating targets from non-targets in GPR data is promising.
Image keypoints are widely used in computer vision for object matching and recognition, where they provide the
best solution for matching and instance recognition of complex objects within cluttered images. Most matching
algorithms operate by rst nding interest points, or keypoints, that are expected to be common across multiple
views of the same object. A small area, or patch, around each keypoint can be represented by a numerical
descriptor that describes the structure of the patch. By matching descriptors from keypoints found in 2-D
data to keypoints of known origin, matching algorithms can determine the likelihood that any particular patch
matches a pre-existing template. The objective in this research is to apply these methods to two-dimensional
slices of Ground Penetrating Radar (GPR) data in order to distinguish between landmine and non-landmine
responses. In this work, a variety of established object matching algorithms have been tested and evaluated
to examine their application to GPR data. In addition, GPR specic keypoint and descriptor methods have
been developed which better suit the landmine detection task within GPR data. These methods improve on the
performance of standard image processing techniques, and show promise for future work involving translations
of technologies from the computer vision eld to landmine detection in GPR data.
KEYWORDS: General packet radio service, LIDAR, Land mines, Detection and tracking algorithms, Antennas, Metals, Target detection, Sensors, Prototyping, Global Positioning System
Vehicle-mounted ground-penetrating radar (GPR) has proved to be a valuable technology for buried threat
detection, especially in the area of military route clearance. However, detection performance may be degraded in
very rough terrain or o-road conditions. This is because the signal processing approaches for target detection
in GPR rst identify the ground re
ection in the data, and then align the data in order to remove the ground
re
ection. Under extremely rough terrain, antenna bounce and multipath eects render nding the ground
re
ection a dicult task, and errors in ground localization can lead to data alignment that distorts potential
target signatures and/or creates false alarms. In this work, commercial-o-the-shelf light detection and ranging
(LIDAR), global positioning system (GPS), and inertial measurement unit (IMU) were integrated with a GPR
into a prototype route clearance system. The LIDAR provided high-resolution measurements of the ground
surface prole, and the GPS/IMU recorded the vehicle's position and orientation. Experiments investigated
the applicability of the integrated system for nding the ground re
ection in GPR data and decoupling vehicle
motion from the rough surface response. Assessment of ground-tracking performance was based on an experiment
involving three prepared test lanes, each with dierent congurations of buried targets and terrain obstacles.
Several algorithms for target detection in GPR were applied to the data, both with traditional preprocessing and
incorporating the LIDAR and IMU. Experimental results suggest that the LIDAR and IMU may be valuable
components for ground tracking in next-generation GPR systems.
Roadside explosive threats continue to pose a significant risk to soldiers and civilians in conflict areas around the world.
These objects are easy to manufacture and procure, but due to their ad hoc nature, they are difficult to reliably detect
using standard sensing technologies. Although large roadside explosive hazards may be difficult to conceal in rural
environments, urban settings provide a much more complicated background where seemingly innocuous objects (e.g.,
piles of trash, roadside debris) may be used to obscure threats. Since direct detection of all innocuous objects would flag
too many objects to be of use, techniques must be employed to reduce the number of alarms generated and highlight only
a limited subset of possibly threatening regions for the user. In this work, change detection techniques are used to
reduce false alarm rates and increase detection capabilities for possible threat identification in urban environments. The
proposed model leverages data from multiple video streams collected over the same regions by first applying video
aligning and then using various distance metrics to detect changes based on image keypoints in the video streams. Data
collected at an urban warfare simulation range at an Eastern US test site was used to evaluate the proposed approach, and
significant reductions in false alarm rates compared to simpler techniques are illustrated.
Many effective buried threat detection systems rely on close proximity and near vertical deployment over subsurface
objects before reasonable performance can be obtained. A forward-looking sensor configuration, where
an object can be detected from much greater distances, allows for safer detection of buried explosive threats,
and increased rates of advance. Forward-looking configurations also provide an additional advantage of yielding
multiple perspectives and looks at each subsurface area, and data from these multiple pose angles can be potentially
exploited for improved detection. This work investigates several aspects of detection algorithms that can
be applied to forward-looking imagery. Previous forward-looking detection algorithms have employed several
anomaly detection algorithms, such as the RX algorithm. In this work the performance of the RX algorithm
is compared to a scale-space approach based on Laplcaian of Gaussian filtering. This work also investigates
methods to combine the detection output from successive frames to aid detection performance. This is done by
exploiting the spatial colocation of detection alarms after they are mapped from image coordinates into world
coordinates. The performance of the resulting algorithms are measured on data from a forward-looking vehicle
mounted optical sensor system collected over several lanes at a western U.S. test facility. Results indicate that
exploiting the spatial colocation of detections made in successive frames can yield improved performance.
Laser induced breakdown spectroscopy (LIBS) can provide rapid, minimally destructive, chemical analysis of
substances with the benefit of little to no sample preparation. Therefore, LIBS is a viable technology for the
detection of substances of interest in near real-time fielded remote sensing scenarios. Of particular interest to
military and security operations is the detection of explosive residues on various surfaces. It has been demonstrated
that LIBS is capable of detecting such residues, however, the surface or substrate on which the residue
is present can alter the observed spectra. Standard chemometric techniques such as principal components analysis
and partial least squares discriminant analysis have previously been applied to explosive residue detection,
however, the classification techniques developed on such data perform best against residue/substrate pairs that
were included in model training but do not perform well when the residue/substrate pairs are not in the training
set. Specifically residues in the training set may not be correctly detected if they are presented on a previously
unseen substrate. In this work, we explicitly model LIBS spectra resulting from the residue and substrate to
attempt to separate the response from each of the two components. This separation process is performed jointly
with classifier design to ensure that the classifier that is developed is able to detect residues of interest without
being confused by variations in the substrates. We demonstrate that the proposed classification algorithm provides
improved robustness to variations in substrate compared to standard chemometric techniques for residue
detection.
The classification of firearms from their acoustic signatures has many potential benefits for a variety of military
and security operations. Most approaches to acoustic gunshot classification can be characterized as frame based
feature classification approaches, where the time-domain acoustic signal is partitioned into a set of frames from
which characterizing features are extracted and used to classify the signals. Although this approach can be
quite successful, performance is highly dependent upon the relationship between the selected frame size and the
signals under consideration. In this work we consider a statistical model for time-domain gunshot signatures
which eliminates the need for both data partitioning and the selection of characterizing features. Each class of
acoustic signals is modeled as a hidden Markov model (HMM) with autoregressive (AR) source densities. Each
AR model specifies a set of spectral and energy characteristics of the signal while the HMM characterizes the
transitions between these states. The model is constructed using nonparametric Bayesian techniques to allow
model inference to learn the number of states within the HMM and the AR order of each state density. The
model thus selects the number of unique spectral components and the complexity of each of these components
from the set of training data, limiting model over-fitting and eliminating the need to optimize performance over
these parameters. We demonstrate that classification using the proposed statistical model performs comparably
to existing techniques without requiring user specified features, thus allowing the same statistical models to be
used on future datasets without modification.
KEYWORDS: General packet radio service, Image segmentation, Land mines, Feature extraction, Image processing algorithms and systems, Detection and tracking algorithms, Data processing, Sensors, Image filtering, Ground penetrating radar
Ground-penetrating radar (GPR) sensors provide an effective means for detecting changes in the sub-surface
electrical properties of soils, such as changes indicative of landmines or other buried threats. However, most
GPR-based pre-screening algorithms only localize target responses along the surface of the earth, and do not
provide information regarding an object's position in depth. As a result, feature extraction algorithms are forced
to process data from entire cubes of data around pre-screener alarms, which can reduce feature fidelity and hamper
performance. In this work, spectral analysis is investigated as a method for locating subsurface anomalies in
GPR data. In particular, a 2-D spatial/frequency decomposition is applied to pre-screener flagged GPR B-scans.
Analysis of these spatial/frequency regions suggests that aspects (e.g. moments, maxima, mode) of the frequency
distribution of GPR energy can be indicative of the presence of target responses. After translating a GPR image
to a function of the spatial/frequency distributions at each pixel, several image segmentation approaches can be
applied to perform segmentation in this new transformed feature space. To illustrate the efficacy of the approach,
a performance comparison between feature processing with and without the image segmentation algorithm is
provided.
KEYWORDS: General packet radio service, Detection and tracking algorithms, Interfaces, Algorithm development, Ground penetrating radar, Land mines, Radar, Target detection, Explosives, Data processing
Rough surfaces present an impediment to the detection of buried threats with ground penetrating radar (GPR). Besides
introducing artifacts in the sub-surface due to rough scattering, very rough or uneven surfaces can make inference of the
location of the ground response from GPR data difficult. Since many algorithms rely on the accurate localization of the
air/ground interface, mistakes in ground location inference can cause significant increases in false alarm rates. Many
different approaches to localizing the ground in a particular A-scan have been proposed, but sharing information across
multiple A-scans to form a realistic, smoothly varying ground response over many spatial locations is a difficult problem
that often requires computationally expensive approaches for adequate solutions. In this work we present an application
of the well-known Viterbi algorithm for accurate localization of the air/ground interface based on hypothesized locations
from multiple nearby A-scans. Our implementation of the Viterbi algorithm enables principled incorporation of prior
information into the ground tracking framework, and provides a solution capable of adapting computational complexity
to the severity of the ground localization problem. Furthermore, the Viterbi algorithm can act as a meta-algorithm,
allowing the use of different A-scan based ground detectors as input, for example. This work illustrates how the Viterbi
algorithm can be incorporated into pre-screening algorithms to provide improved target detection rates at lower false
alarm rates.
KEYWORDS: General packet radio service, Data modeling, Land mines, Detection and tracking algorithms, Video processing, Expectation maximization algorithms, Radar, Video, Antennas, Ground penetrating radar
Due to the large amount of data generated by vehicle-mounted ground penetrating radar (GPR) antennae arrays,
advanced feature extraction and classification can only be performed on a small subset of data during real-time
operation. As a result, most GPR based landmine detection systems implement "pre-screening" algorithms to processes
all of the data generated by the antennae array and identify locations with anomalous signatures for more advanced
processing. These pre-screening algorithms must be computationally efficient and obtain high probability of detection,
but can permit a false alarm rate which might be higher than the total system requirements. Many approaches to prescreening
have previously been proposed, including linear prediction coefficients, the LMS algorithm, and CFAR-based
approaches. Similar pre-screening techniques have also been developed in the field of video processing to identify
anomalous behavior or anomalous objects. One such algorithm, an online k-means approximation to an adaptive
Gaussian mixture model (GMM), is particularly well-suited to application for pre-screening in GPR data due to its
computational efficiency, non-linear nature, and relevance of the logic underlying the algorithm to GPR processing. In
this work we explore the application of an adaptive GMM-based approach for anomaly detection from the video
processing literature to pre-screening in GPR data. Results with the ARA Nemesis landmine detection system
demonstrate significant pre-screening performance improvements compared to alternative approaches, and indicate that
the proposed algorithm is a complimentary technique to existing methods.
KEYWORDS: General packet radio service, Data modeling, Land mines, Detection and tracking algorithms, Process modeling, Performance modeling, Ground penetrating radar, Feature selection, 3D modeling, Data processing
In landmine detection applications, fluctuation of environmental and operating conditions can limit the performance
of sensors based on ground-penetrating radar (GPR) technology. As these conditions vary, the classification
and fusion rules necessary for achieving high detection and low false alarm rates may change. Therefore,
context-dependent learning algorithms that exploit contextual variations of GPR data to alter decision rules have
been considered for improving the performance of landmine detection systems. Past approaches to contextual
learning have used both generative and discriminative methods to learn a probabilistic mixture of contexts, such
as a Gaussian mixture, fuzzy c-means clustering, or a mixture of random sets. However, in these approaches the
number of mixture components is pre-defined, which could be problematic if the number of contexts in a data
collection is unknown a priori. In this work, a generative context model is proposed which requires no a priori
knowledge in the number of mixture components. This was achieved through modeling the contextual distribution
in a physics-based feature space with a Gaussian mixture, while also incorporating a Dirichlet process prior
to model uncertainty in the number of mixture components. This Dirichlet process Gaussian mixture model
(DPGMM) was then incorporated in the previously-developed Context-Dependent Feature Selection (CDFS)
framework for fusion of multiple landmine detection algorithms. Experimental results suggest that when the
DPGMM was incorporated into CDFS, the degree of performance improvement over conventional fusion was
greater than when a conventional fixed-order context model was used.
KEYWORDS: Video, General packet radio service, Roads, Land mines, Cameras, Video processing, Explosives, Feature extraction, Ground penetrating radar, Sensors
Forward looking video can provide a large amount of tactically-relevant information to vehicle operators regarding
roadside explosive threats. However it is difficult for vehicle operators to keep track of what roadside objects have
changed since their last excursion, or what new objects have appeared on a road and might therefore contain an
explosive threat. Furthermore, the large amount of data generated by forward looking video can overwhelm users. It
would be of benefit to vehicle operators if only objects that had significantly changed since a recent excursion were
flagged and presented to the user. In this work we develop techniques for video and ground-penetrating radar (GPR)
tracking and aligning, and novel-object identification for application in route clearance patrols. We focus on aligning
video data collected using vehicle mounted forward-looking video cameras and downward-looking GPR using locallyinvariant
feature transforms and set-based distance metrics. Based on these aligned image streams, we then apply
pattern classification approaches to discriminate new explosive threats from stationary and persistent objects. The
techniques described in this work are widely applicable to other forward and downward-looking sensor systems, and are
computationally tractable. The results indicate the potential to robustly identify recently changed roadside threats, and to
present a significantly reduced amount of information to end-users for further operational analysis.
It has been established throughout the ground-penetrating radar (GPR) literature that environmental factors
can severely impact the performance of GPR sensors in landmine detection applications. Over the years, electromagnetic
inversion techniques have been proposed for determining these factors with the goal of mitigating
performance losses. However, these techniques are often computationally expensive and require models and responses
from canonical targets, and therefore may not be appropriate for real-time route-clearance applications.
An alternative technique for mitigating performance changes due to environmental factors is context-dependent
classification, in which decision rules are adjusted based on contextual shifts identified from the GPR data. However,
analysis of the performance of context-dependent learning has been limited to qualitative comparisons of
contextually-similar GPR signatures and quantitative improvement to the ROC curve, while the actual information
extracted regarding soils has not been investigated thoroughly. In this work, physics-based features of GPR
data used in previous context-dependent approaches were extracted from simulated GPR data generated through
Finite-Difference Time-Domain (FDTD) modeling. Statistical techniques where then used to predict several potential
contextual factors, including soil dielectric constant, surface roughness, amount of subsurface clutter,
and the existence of subsurface layering, based on the features. Results suggest that physics-based features of
the GPR background may contain informatin regarding physical properties of the environment, and contextdependent
classification based on these features can exploit information regarding these potentially-important
environmental factors.
KEYWORDS: General packet radio service, Electromagnetic coupling, Land mines, Sensors, Feature extraction, Lawrencium, Target detection, Data fusion, Algorithm development, Antennas
Despite advances in both electromagnetic induction (EMI) and ground penetrating radar (GPR) sensing and related
signal processing, neither sensor alone provides a perfect tool for detecting the myriad of possible buried objects that
threaten the lives of Soldiers and civilians. However, while neither GPR nor EMI sensing alone can provide optimal
detection across all target types, the two approaches are highly complementary. As a result, many landmine systems
seek to make use of both sensing modalities simultaneously and fuse the results from both sensors to improve detection
performance for targets with widely varying metal content and GPR responses. Despite this, little work has focused on
large-scale comparisons of different approaches to sensor fusion and machine learning for combining data from these
highly orthogonal phenomenologies. In this work we explore a wide array of pattern recognition techniques for
algorithm development and sensor fusion. Results with the ARA Nemesis landmine detection system suggest that nonlinear
and non-parametric classification algorithms provide significant performance benefits for single-sensor algorithm
development, and that fusion of multiple algorithms can be performed satisfactorily using basic parametric approaches,
such as logistic discriminant classification, for the targets under consideration in our data sets.
Ground Penetrating Radar (GPR) data provides a powerful technique to identify subsurface buried threats.
Although GPR data contains a three-dimensional representation of the subsurface, object truth (i.e. labels and
positions of true threat objects in training lanes) is often provided in only two dimensions (GPS coordinates
along the earth's surface). To mitigate uncertainty in an object's location in depth, many successful feature extraction/
object recognition techniques in GPR extract feature vectors from several depth regions, and attempt
to combine information across these feature vectors to make final decisions. However, many machine learning
techniques are not well suited for learning under these conditions. Multiple Instance Learning (MIL) is a type of
supervised learning method in which labels are available for sets of samples, but not for individual samples. The
goal of learning in MIL is to classify new sets of samples as they become available. This set-based framework
is useful in processing GPR responses since features are often extracted independently from multiple un-labeled
depth bins, and thus a set of features is produced at each potential threat location. In this work, a comparison
of several previous approaches to MlL applied to landmine detection in GPR data is presented. One recent
algorithm, the p-Posterior Mixture Model approach (pPMM) is given special attention, and several slight modifications
to the pPMM approach are presented and compared.
KEYWORDS: Detection and tracking algorithms, Land mines, Antennas, General packet radio service, Ground penetrating radar, Interfaces, Algorithm development, Reflection, Sensors, Data corrections
In using GPR images for landmine detection it is often useful to identify the air-ground interface in the GPR
signal for alignment purposes. A number of algorithms have been proposed to solve the air-ground interface
detection problem, including some which use only A-scan data, and others which track the ground in B-scans or
C-scans. Here we develop a framework for comparing these algorithms relative to one another and we examine
the results. The evaluations are performed on data that have been categorized in terms of features that make the
air-ground interface difficult to find or track. The data also have associated human selected ground locations,
from multiple evaluators, that can be used for determining correctness. A distribution is placed over each of the
human selected ground locations, with the sum of these distributions at the algorithm selected location used as
a measure of its correctness. Algorithms are also evaluated in terms of how they affect the false alarm and true
positive rates of mine detection algorithms that use ground aligned data.
KEYWORDS: Data modeling, Sensors, Land mines, Signal to noise ratio, Target detection, Mining, Electromagnetic coupling, Distance measurement, Model-based design, Electromagnetism
Frequency-domain electromagnetic induction (EMI) sensors have the ability to provide target signatures which enable
discrimination of landmines from harmless clutter. In particular, frequency-domain EMI sensors are well-suited for target
characterization by inverting a physics-based signal model. In many model-based signal processing paradigms, the
target signatures can be decomposed into a weighted sum of parameterized basis functions, where the basis functions
are intrinsic to the target under consideration and the associated weights are a function of the target sensor orientation.
The basis function parameters can then be used as features for classification of the target as landmine or clutter. In this
work, frequency-domain EMI sensor data feature extraction and processing is investigated, with a variety of physics-based
models and statistical classifiers considered. Results for data measured with a prototype frequency-domain EMI sensor
at a standardized test site are presented. Preliminary results indicate that extracting physics-based features followed by
statistical classification provides an effective approach for classifying targets as landmine or clutter.
KEYWORDS: General packet radio service, Land mines, Feature extraction, Data modeling, Autoregressive models, Feature selection, Fourier transforms, Detection and tracking algorithms, Soil science, Reflection
Context-dependent classification techniques applied to landmine detection with ground-penetrating radar (GPR)
have demonstrated substantial performance improvements over conventional classification algorithms. Context-dependent
algorithms compute a decision statistic by integrating over uncertainty in the unknown, but probabilistically
inferable, context of the observation. When applied to GPR, contexts may be defined by differences
in electromagnetic properties of the subsurface environment, which are due to discrepancies in soil composition,
moisture levels, and surface texture. Context-dependent Feature Selection (CDFS) is a technique developed for
selecting a unique subset of features for classifying landmines from clutter in different environmental contexts.
In past work, context definitions were assumed to be soil moisture conditions which were known during training.
However, knowledge of environmental conditions could be difficult to obtain in the field. In this paper, we utilize
an unsupervised learning algorithm for defining contexts which are unknown a priori. Our method performs unsupervised
context identification based on similarities in physics-based and statistical features that characterize
the subsurface environment of the raw GPR data. Results indicate that utilizing this contextual information
improves classification performance, and provides performance improvements over non-context-dependent approaches.
Implications for on-line context identification will be suggested as a possible avenue for future work.
Time domain ground penetrating radar (GPR) has been shown to be a powerful sensing phenomenology for
detecting buried objects such as landmines. Landmine detection with GPR data typically utilizes a feature-based
pattern classification algorithm to discriminate buried landmines from other sub-surface objects. In high-fidelity
GPR, the time-frequency characteristics of a landmine response should be indicative of the physical
construction and material composition of the landmine and could therefore be useful for discrimination from
other non-threatening sub-surface objects. In this research we propose modeling landmine time-domain responses
with a nonparametric Bayesian time-series model and we perform clustering of these time-series models with a
hierarchical nonparametric Bayesian model. Each time-series is modeled as a hidden Markov model (HMM) with
autoregressive (AR) state densities. The proposed nonparametric Bayesian prior allows for automated learning
of the number of states in the HMM as well as the AR order within each state density. This creates a flexible
time-series model with complexity determined by the data. Furthermore, a hierarchical non-parametric Bayesian
prior is used to group landmine responses with similar HMM model parameters, thus learning the number of
distinct landmine response models within a data set. Model inference is accomplished using a fast variational
mean field approximation that can be implemented for on-line learning.
Ground penetrating radar (GPR) and electromagnetic induction (EMI) sensors provide complementary capabilities in
detecting buried targets such as landmines, suggesting that the fusion of GPR and EMI modalities may provide
improved detection performance over that obtained using only a single modality. This paper considers both pre-screening
and the discrimination of landmines from non-landmine objects using real landmine data collected from a
U.S. government test site as part of the Autonomous Mine Detection System (AMDS) landmine program. GPR and
EMI pre-screeners are first reviewed and then a fusion pre-screener is presented that combines the GPR and EMI prescreeners
using a distance-based likelihood ratio test (DLRT) classifier to produce a fused confidence for each pre-screener
alarm. The fused pre-screener is demonstrated to provide substantially improved performance over the
individual GPR and EMI pre-screeners.
The discrimination of landmines from non-landmine objects using feature-based classifiers is also considered. The
GPR feature utilized is a pre-processed, spatially filtered normalized energy metric. Features used for the EMI sensor
include model-based features generated from the AETC model and a dipole model as well as features from a matched
subspace detector. The EMI and GPR features are then fused using a random forest classifier. The fused classifier
performance is superior to the performance of classifiers using GPR or EMI features alone, again indicating that
performance improvements may be obtained through the fusion of GPR and EMI sensors. The performance
improvements obtained both for pre-screening and for discrimination have been verified by blind test results scored by
an independent U.S. government contractor.
Frequency-domain electromagnetic induction (EMI) sensors have the ability to measure target signatures which enable
discrimination of landmines from harmless clutter. In a model-based signal processing paradigm, the target signatures can
be decomposed into a weighted sum of parameterized basis functions, where the basis functions are intrinsic to the target
under consideration and the associated weights are a function of the target sensor orientation. The basis function parameters
can then be used as features for classification of the target as landmine or clutter. One of the challenges associated with
effectively utilizing frequency-domain EMI sensor data within a model-based signal processing paradigm such as this is
determining the correct model order for the measured data, as the number of basis functions intrinsic to the target under
consideration is not known a priori. In this work, relevance vector machine (RVM) regression is applied to simultaneously
determine both the number of parameterized basis functions and their relative contributions to the measured signal. The
target may then be classified utilizing the basis function parameters as features within a statistical classifier. Results for
data measured with a prototype frequency-domain EMI sensor at a standardized test site are presented. Preliminary results
indicate that RVM regression followed by statistical classification utilizing the resulting model-based features provides an
effective approach for classifying targets as landmine or clutter.
KEYWORDS: General packet radio service, Autoregressive models, Land mines, Feature extraction, Feature selection, Detection and tracking algorithms, Environmental sensing, Data modeling, Antennas, Ground penetrating radar
We present a novel method for improving landmine detection with ground-penetrating radar (GPR) by utilizing
a priori knowledge of environmental conditions to facilitate algorithm training. The goal of Context-Dependent
Feature Selection (CDFS) is to mitigate performance degradation caused by environmental factors. CDFS
operates on GPR data by first identifying its environmental context, and then fuses the decisions of several
classifiers trained on context-dependent subsets of features. CDFS was evaluated on GPR data collected at several
distinct sites under a variety of weather conditions. Results show that using prior environmental knowledge in
this fashion has the potential to improve landmine detection.
Recent advances in Laser-Induced breakdown spectroscopy (LIBS), Raman spectroscopy, and other spectroscopic
approaches have increased interest in the application of spectroscopy to detection of explosives along with other
chemical-signature identification tasks. However most existing spectroscopic data collection techniques require
manual interaction with data files including data manipulation using multiple pieces of software and different
file formats, time-consuming feature-selection, and model re-generation. Not only do these steps reduce analytic
efficiency and slow the progress of research in spectroscopy, but they also inhibit real-time use of the systems by
end-users. In this work we present a graphical user interface designed to increase efficiency for spectroscopic data
collection, feature selection, classifier development, and testing. We present a software architecture that provides
enough flexibility to handle data from many different spectroscopic sensors. We also discuss feature-level and
model-level software components that allow for the features and classification approaches to be manipulated
interactively, and we present a simple and intuitive testing screen suitable for an end user to make decisions in
the field with out requiring a "human in the loop" for processing.
Current ground penetrating radar algorithms for landmine detection require accurate estimates of the location
of the air/ground interface to maintain high levels of performance. However, the presence of surface clutter,
natural soil roughness, and antenna motion lead to uncertainty in these estimates. Previous work on improving
estimates of the location of the air/ground interface have focused on one-dimensional filtering techniques to
localize the air/ground interface. In this work, we propose an algorithm for interface localization using a 2-
D Gaussian Markov random field (GMRF). The GMRF provides a statistical model of the surface structure,
which enables the application of statistical optimization techniques. In this work, the ground location is inferred
using iterated conditional modes (ICM) optimization which maximizes the conditional pseudo-likelihood of the
GMRF at a point, conditioned on its neighbors. To illustrate the efficacy of the proposed interface localization
approach, pre-screener performance with and without the proposed ground localization algorithm is compared.
We show that accurate localization of the air/ground interface provides the potential for future performance
improvements.
KEYWORDS: Sensors, Land mines, Electromagnetic coupling, Sensor performance, Electromagnetism, Detection and tracking algorithms, Metals, Mining, Platinum, Signal to noise ratio
This work provides a performance comparison between two
frequency-domain electromagnetic induction (EMI)
sensors - one quadrupole and one dipole sensor for the detection of subsurface anti-personnel and anti-tank
landmines. A summary of the physical differences between the two sensors and from those of other EMI sensors
will be discussed. Previously we presented a performance analysis of the dipole sensor for a variety of detection
algorithms over data collected at a government test facility indicating robust performance using the dipole
sensor. The algorithms considered previously included an energy detector, matched subspace detector and a
kNN probability density estimation approach over the features of a four parameter phenomenological model.
The current sensor comparison will include, in addition to the previous detection methods, a Random Forests
classification algorithm and utilize a larger training data set.
Recent advances in ground penetrating radar (GPR) design and fabrication have resulted in improved fidelity
responses from relatively small, shallow-buried objects like landmines and improvised explosive devices. As the
responses measured with GPR improve, more and more advanced processing techniques can be brought to bear
on the problem of target identification in GPR data. From an electromagnetic point of view, the problem of
target detection in GPR signal processing is reducible to inferring the presence or absence of changes in the
electromagnetic properties of soils and thus the presence or absence of buried targets. Problems arise because
the algorithms required for the full electromagnetic inversion of GPR signals are extremely computationally
expensive, and usually rely on assumptions of electromagnetically constant transmission media; these problems
typically make the real-time implementation of purely electromagnetic-inspired algorithms infeasible. On the
other hand, purely statistical or signal-processing inspired approaches to target identification in GPR often lack a
solid theoretical basis in the underlying physics, which is fundamental to understanding responses in GPR. In this
work, we propose a model for responses in time-domain ground penetrating radar that attempts to incorporate
the underlying physics of the problem, but avoids several of the issues inherent in assuming constant media with
known electrical parameters by imposing a statistical model over the observed parameters of interest in A-scans - namely the signal gains, times of arrival, etc. The spatial requirements of the proposed statistical model suggests
the application of Markov random field (MRF) distributions which provide expressive, but computationally
simple models of spatial interactions. In this work we will explore the application of physics-based MRF's
as generative models for time-domain GPR data, the pre-screening algorithms that this model motivates, and
discuss how the model can be extended to other applications in GPR processing. Preliminary results showing
how the MRF approach to understanding the underlying physics can improve performance are also shown.
As ground penetrating radar sensor phenomenology improves, more advanced statistical processing approaches
become applicable to the problem of landmine detection in GPR data. Most previous studies on landmine
detection in GPR data have focused on the application of statistics and physics based prescreening algorithms,
new feature extraction approaches, and improved feature classification techniques. In the typical framework,
prescreening algorithms provide spatial location information of anomalous responses in down-track / cross-track
coordinates, and feature extraction algorithms are then tasked with generating low-dimensional information-bearing
feature sets from these spatial locations. However in time-domain GPR, a significant portion of the data
collected at prescreener flagged locations may be unrelated to the true anomaly responses - e.g. ground bounce
response, responses either temporally "before" or "after" the anomalous response, etc. The ability to segment
the information-bearing region of the GPR image from the background of the image may thus provide improved
performance for feature-based processing of anomaly responses. In this work we will explore the application of
Markov random fields (MRFs) to the problem of anomaly/background segmentation in GPR data. Preliminary
results suggest the potential for improved feature extraction and overall performance gains via application of
image segmentation approaches prior to feature extraction.
KEYWORDS: Land mines, General packet radio service, Detection and tracking algorithms, Data modeling, Filtering (signal processing), Radar, Error analysis, Statistical modeling, Sensors, Systems modeling
A Sequential Monte Carlo (SMC) method is proposed to locate the ground bounce (GB) positions in 3D data collected by ground penetrating radar (GPR) system. The algorithm is verified utilizing real data and improved landmine detection performance is achieved compared with three other GB trackers.
KEYWORDS: Sensors, Land mines, Data modeling, Electromagnetic coupling, Mining, Electromagnetism, Systems modeling, Performance modeling, Detection and tracking algorithms, Feature extraction
This work explores possible performance enhancements for landmine detection algorithms using frequency domain
wideband electromagnetic induction sensors. A pre-existing four parameter model for conducting objects
based on empirically collected data for UXO is discussed, and its application for accurately modeling landmine
signatures is also considered. Discrimination of mines versus clutter based on the extracted model parameters is
considered. Furthermore, this work will compare the effectiveness of discrimination based on the four parameter
model to a matched subspace detection algorithm. Experimental results using data from government run test sites will be presented.
In this work we explore and compare several statistical pattern recognition techniques for classification and identification of buried landmines using both electromagnetic induction and ground penetrating radar data. In particular we explore application of different feature extraction approaches to the problem of landmine/clutter classification in blind- and known- ground truth scenarios using data from the NIITEK ground penetrating radar and the Vallon EMI sensor as well as the CyTerra GPR and Minelab EMI sensors. We also compare and contrast the generalization capabilities of different kernels including radial basis function, linear, and direct kernels within the relevance vector machine framework. Results are presented for blind-test scenarios that illustrate robust classification for features that can be extracted with low computational complexity.
KEYWORDS: Land mines, Autoregressive models, General packet radio service, Digital filtering, Genetic algorithms, Data modeling, Signal processing, Electronic filtering, Detection and tracking algorithms, Optimization (mathematics)
Previous large-scale blind tests of anti-tank landmine detection utilizing the NIITEK ground penetrating radar indicated the potential for very high anti-tank landmine detection probabilities at very low false alarm rates for algorithms based on adaptive background cancellation schemes. Recent data collections under more heterogeneous multi-layered road-scenarios seem to indicate that although adaptive solutions to background cancellation are effective, the adaptive solutions to background cancellation under different road conditions can differ significantly, and misapplication of these adaptive solutions can reduce landmine detection performance in terms of PD/FAR. In this work we present a framework for the constrained optimization of background-estimation
filters that specifically seeks to optimize PD/FAR performance as measured by the area under the ROC curve between two FARs. We also consider the application of genetic algorithms to the problem of filter optimization for landmine detection. Results indicate robust results for both static and adaptive background cancellation schemes, and possible real-world advantages and disadvantages of static and adaptive approaches are discussed.
KEYWORDS: Sensors, Target detection, Land mines, Ground penetrating radar, Radar, General packet radio service, Signal detection, Detection and tracking algorithms, Target recognition, Data modeling
In this work we present an application of matched subspace detectors to the problem of target detection and identification using ground penetrating radar data. In particular we apply sets of matched subspace detector filter banks to data containing both anti-personnel and anti-tank targets as well as metallic and non-metallic clutter objects. Current results indicate the potential for robust target detection and identification but further improvements via subspace modeling and signal extraction/enhancement may also improve performance.
KEYWORDS: Sensors, Target detection, Electromagnetic coupling, Metals, Land mines, Data acquisition, Mining, Calibration, Prototyping, Signal to noise ratio
Last year, we reported on a preliminary evaluation of GE’s frequency-domain EMI prototype sensor capable of measuring the wideband response of simulant and inert low metal mines at shallow depths over a frequency range from 100 Hz to 150 kHz. Since then, the prototype sensor has undergone further power and sensitivity improvements and has been taken to the field to collect signature data on targets in a calibration grid located at an Army facility in Virginia. The frequency-domain EMI responses have been analyzed by Duke University utilizing matched subspace detector (MSD) processing. The limited amount of data collected, so far, suggests that MSD processing of the frequency-domain data is a robust technique for target detection and identification. However, more data need to be collected for robust testing.
KEYWORDS: Sensors, Land mines, Sensor performance, Active sensors, Target detection, Detection and tracking algorithms, General packet radio service, Electromagnetic coupling, Binary data, Computer simulations
We consider an information-theoretic approach for sensor management that chooses sensors and sensor parameters in order to maximize the expected discrimination gain associated with each new sensor measurement. We analyze the problem of searching for N targets with M multimodal sensors, where each sensor has its own probability of detection, probability of false alarm, and cost of use. Other information, such as the prior distribution of the targets in space and the degree of constraint of the sensor motion, is also utilized in our formulation. Performance of the sensor management algorithm is then compared to the performance of a direct-search procedure in which the sensors blindly search through all cells in a predetermined path. The information-based sensor manager is found to have significant performance gains over the direct-search approach. Algorithm performance is also analyzed using real landmine data taken with three different sensing modalities. Detection performance using the sensor management algorithm is again found to be superior to detection performance using a blind search procedure. The simulation and real-data results also both illuminate the increased performance available through multimodal sensing.
Recently, blind tests of several automated detection algorithms operating on the NIITEK ground penetrating radar data (GPR) have resulted in quite promising performance results. Anecdotally, human observers have also shown notable skill in detecting landmines and rejecting false alarms in this same data; however, the basis of human performance has not been studied in depth. In this study, human observers are recruited from the undergraduate and graduate student population at Duke University and are trained to visually detect landmines in the NIITEK GPR data. Subjects are then presented with GPR responses associated with blanks, clutter items (including emplaced clutter), and landmines in a blind test scenario. Subjects are asked to make the decision as to whether they are viewing a landmine response or a false alarm, and their performance is scored. A variety of landmines, measured at several test sites, are presented to determine the relative difficulty in detecting each mine type. Subject performance is compared to the performance of two automated algorithms already under development for the NIITEK radar system: LMS and FROSAW. In addition, subjects are given a subset of features for each alarm from which they may indicate the reason behind their decision. These last data may provide a basis for the design of an automated algorithm that takes advantage of the most useful of the observed features.
With current signal processing techniques, successful discrimination between UXO (Unexploded Ordnance) and clutter depends on characteristics that are consistent across all examples of an ordnance type. Real UXO, however, exhibit many differences from instance to instance, such as varying degrees of damage sustained, degradation over time, orientation in the ground, and even differences in design. Thus, a given ordnance type, such as 60mm shells, will exhibit a wide range of signal responses, making it difficult to distinguish these items from clutter unless these fundamental differences are taken into account using appropriate mathematics. This paper will examine optimal methods of using frequency-domain analysis of wideband electromagnetic induction (EMI) for detection.
KEYWORDS: Mining, General packet radio service, Ground penetrating radar, Land mines, Antennas, 3D modeling, Reflectors, Data modeling, Radar, Dielectrics
Two new features are presented to improve the detection of Anti-Tank (AT) landmines using Ground Penetrating Radar (GPR). A simplified three dimensial physics based model is used as the basis for the features. We combine these features with the results of an algorithm known as LMS. We present promising feature detection algorithms known as Rings N' Things (RNT) and Cross Diagonal Enhancement Processing (CDEP) and our approach to combining the new features with the LMS features using logistic regression techniques. Test results from data gathered at multiple sites covering hundreds of mines and thousands of square meters is analyzed and presented.
KEYWORDS: Sensors, General packet radio service, Metals, Land mines, Detection and tracking algorithms, Electromagnetic coupling, Mining, Sensor fusion, Algorithm development, Fusion energy
The recent development of high quality sensors paired with development of advanced statistical signal processing algorithms has shown that there are sensors that can not only discriminate targets from clutter, but can also identify subsurface or obscured targets. In a previous theoretical and simulation study, we utilized this identification capability in addition to contextual information in a multi-modal adaptive algorithm where the identification capabilities of multiple sensors are utilized to modify the prior probability density functions associated with statistical models being utilized by other sensors. We assumed that the statistics describing the features associated with each sensor modality follow a Gaussian mixture density, where in many cases the individual Gaussian distributions that make up the mixture result from different target types or target classes. We utilized identification information from one sensor to modify the weights associated with the probability density functions being utilized by algorithms associated with other sensor modalities. In our simulations, this approach is shown to be improve sensor performance by reducing the overall false alarm rate. In this talk, we transition the approach from a simulation study to consider real field data collected by both handheld and vehicular based systems. We show that by appropriate modification of our statistical models to accurately match field data, improved performance can be obtained over traditional sensor fusion algorithms.
In this paper we present a multi-stage algorithm for target/clutter discrimination and target identification using the Niitek/Wichmann ground penetrating radar (GPR). To identify small subsets of GPR data for feature-processing, a pre-screening algorithm based on the 2-D lattice least mean squares (LMS) algorithm is used to flag locations of interest. Features of the measured GPR data at these flagged locations are then generated and pattern recognition techniques are used to identify targets using these feature sets. It has been observed that trained human subjects are often quite successful at discriminating targets from clutter. Some features are designed to take advantage of the visual aberrations that a human observer might use. Other features based on a variety of image and signal processing techniques are also considered. Results presented indicate improvements for feature-based processors over pre-screener algorithms.
KEYWORDS: Detection and tracking algorithms, Ground penetrating radar, Radar, General packet radio service, Global Positioning System, Land mines, Metals, Digital filtering, Sensors, Image processing algorithms and systems
This paper describes the application of a 2-dimensional (2-D) lattice LMS algorithm for anomaly detection using the Wichmann/Niitek ground penetrating radar (GPR) system. Sets of 3-dimensional (3-D) data are collected from the GPR system and these are processed in separate 2-D slices. Those 2-D slices that are spatially correlated in depth are combined into separate “depth segments” and these are processed independently. When target/no target declarations need to be made, the individual depth segments are combined to yield a 2-D confidence map. The 2-D confidence map is then thresholded and alarms are placed at the centroids of the remaining 8-connected data points. Calibration lane results are presented for data collected over several soil types under several weather conditions. Results show a false alarm rate improvement of at least an order of magnitude over other GPR systems, as well as significant improvement over other adaptive algorithms operating on the same data.
Wideband electromagnetic induction (EMI) data provides an opportunity to apply statistical signal processing techniques to potentially mitigate false alarm rates in landmine detection. This paper explores the application of matched subspace detectors and support vector machines (SVMs) to this problem. A library of landmine responses is generated from background-corrected calibration data and a bank of matched subspace detectors, each tuned to a specific mine type, is generated. Support vector machines are implemented based on the full mine responses, decay rate estimates, and the outputs of the matched subspace filter banks. Different training approaches are considered for the support vector machines. Receiver operating characteristics (ROCs) for the matched subspace detectors and support vector machines operating in a blind field test are presented. The results indicate that substantial reductions in the false alarm rates can be achieved using these techniques.
Ground penetrating radar has been proposed as an alternative sensor to classical electromagnetic induction techniques for the landmine detection problem. The NIITEK-Wichmann antenna provides a high frequency radar signal with very low noise levels following the ground reflection. As a result, the signal from a buried object is not masked by the inherent noise in the system. It has been demonstrated that an operator can learn to interpret the NIITEK-Wichmann radar signal to detect and identify buried targets. The goal of this work is to develop signal processing algorithms to automatically process the radar signals and differentiate between targets and clutter. The algorithms that we are investigating have been tested on data collected at the JUXOCO test grid as well as on data collected in calibration lanes that are used for evaluating the performance of handheld and vehicular landmine detection systems. We have developed algorithms based on principle component analysis, independent component analysis, matched filters, and Bayesian processing of wavelet features. We have also considered several approaches to ground-bounce removal prior to processing. In this paper we discuss the relative performance of each of the techniques as well as the impact of ground bounce removal on processing of the data.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.