Show all abstracts
View Session
- Front Matter: Volume 8017
- Electromagnetic Induction I
- Electromagnetic Induction II
- A Mélange of Techniques
- Sensing and Detecting in the Marine Environment I
- Sensing and Detecting in the Marine Environment II
- Soils and Soil Effects
- Detection of Bulk Explosives Threats I
- Detection of Bulk Explosives Threats II
- Standoff Detection of Explosives
- Advances in Ground Penetrating Radar Subsurface Object Detection
- Tracking Rough Ground in Ground-Penetrating Radar Data
- Signal Processing Ground-Penetrating Radar Data I
- Signal Processing Ground-Penetrating Radar Data II
- Infrared
- Signal Processing and Sensor Fusion
Front Matter: Volume 8017
Front Matter: Volume 8017
Show abstract
This PDF file contains the front matter associated with SPIE Proceedings Volume 8017, including the Title Page, Copyright information, Table of Contents, Introduction, and the Conference Committee listing.
Electromagnetic Induction I
Open-area concealed-weapon detection system
Show abstract
Concealed Weapon Detection (CWD) has become a significant challenge to present day security needs; individuals
carrying weapons into airplanes, schools, and secured establishments are threat to public security. Although controlled
screening, of people for concealed weapons, has been employed in many establishments, procedures and equipment are
designed to work in restricted environments like airport passport control, military checkpoints, hospitals, school and
university entrance. Furthermore, screening systems do not effectively decipher between threat and non-threat metal
objects, thus leading to high rate of false alarms which can become a liability to daily operational needs of
establishments. Therefore, the design and development of a new CWD system to operate in a large open area
environment with large numbers of people reduced incidences of false alarms and increased location accuracy is
essential.
Magnetic sensing techniques for humanitarian ordnance detection and discrimination
Show abstract
Detection and discrimination of unexploded ordnance (UXO) in areas of prior conflict is of high importance to the
international community and the United States government. For humanitarian applications, sensors and processing
methods need to be robust, reliable, and easy to train and implement using indigenous UXO removal personnel. This
paper focuses on magnetometer sensing techniques, processing, and operation for UXO detection and discrimination
applications. Specifically, we discuss the collection, processing, and discrimination of data collected using the PACMAG
man-portable system consisting of arrays of sensitive total-field magnetometers, global positioning (GPS)
combined with digital odometers, and a data acquisition system. We outline preliminary standard operating procedures
for optimal collection of UXO-induced magnetic fields and associated position data using either a GPS, or odometer
when surveying in GPS-denied areas. Processing techniques such as gridding and filtering, target picking, and
discrimination lead to estimates of target size and location. Emphasis is placed on simplifying the production of
magnetometer hardware and software for use by minimally-trained personnel with no advanced knowledge of magnetic
sensing and geophysics.
Incorporating advanced EMI technologies in operational munitions characterization surveys
Show abstract
The presence of unexploded ordnance (UXO), discarded military munitions (DMM), and munitions constituents (MC) at
both active and formerly used defense sites (FUDS) has created a necessity for production-level efforts to remove these
munitions and explosives of concern (MEC). Ordnance and explosives (OE) and UXO removal operations typically
employ electromagnetic induction (EMI) or magnetometer surveys to identify potential MEC hazards in previously
determined areas of interest. A major cost factor in these operations is the significant allocation of resources for the
excavation of harmless objects associated with fragmentation, scrap, or geological clutter. Recent advances in
classification and discrimination methodologies, as well as the development of sensor technologies that fully exploit
physics-based analysis, have demonstrated promise for significantly reducing the false alarm rate due to MEC related
clutter. This paper identifies some of the considerations for and the challenges associated with implementing these
discrimination methodologies and advanced sensor technologies in production-level surveys. Specifically, we evaluate
the implications of deploying an advanced multi-axis EMI sensor at a variety of MEC sites, the discrimination
methodologies that leverage the data produced by this sensor, and the potential for productivity increase that could be
realized by incorporating this advanced technology as part of production protocol.
Rapid position estimation using electromagnetic induction data from the MetalMapper in dynamic mode
Show abstract
Dynamic data from the MetalMapper electromagnetic induction sensor are analyzed using a fast inversion algorithm
in order to obtain position information of buried anomalies. After validating the algorithm by comparing
static and dynamic inversions from reference measurements at Camp San Luis Obispo, the algorithm is applied
to realistic dynamic measurements from Camp Butner. A sequence of 939 data points are inverted as the
MetalMapper travels along a calibration lane, flagging a few positions as corresponding to buried anomalies. An
a posteriori comparison with field plots reveals a good agreement between the flagged positions and the field
peak values, suggesting the efficacy of the algorithm at detecting a large variety of anomalies from dynamic
data.
Comparison of supervised and unsupervised machine learning techniques for UXO classification using EMI data
Show abstract
Classification tools including Support Vector Machines (SVM) and Neural Networks (NN) are employed, and their
performances compared for Unexploded Ordnance (UXO) classification using live site electromagnetic induction (EMI)
data. Both SVM and NN are examples of supervised machine-learning techniques, whose purpose is to label the features
(extracted from the incoming data of the unknown subsurface anomalies) based on previously trained examples. In this
paper a set of three features are extracted from the EMI decay curves of the physics-based intrinsic, effective dipole
moment, called the total Normalized Surface Magnetic Source (NSMS). This data is first used to train both the SVM and
NN models and further serves as a basis for UXO classification. These techniques are then compared to an unsupervised
learning approach, based on agglomerative hierarchical clustering followed by Gaussian Mixture modeling. We found
that such combination provides reduction in the amount of required training data (which is being requested solely based
on the clustering results) and allows for convenient probabilistic interpretation of the classification. The classification
results themselves depend on the UXO caliber, material composition and actual live UXO site's conditions. Therefore,
here we report the classification results for a live UXO data set, collected at former Camp San Luis Obispo, CA. This
study includes four targets-of-interest: 60-mm, 81-mm, and 4.2-in mortars and 2.36-in rockets. The classification
performance between clutters and UXO is studied and the corresponding ROC curves are illustrated.
MPV-II: an enhanced vector man-portable EMI sensor for UXO identification
Show abstract
The Man-Portable Vector (MPV) electromagnetic induction sensor has proved its worth and flexibility as a tool for identification
and discrimination of unexploded ordnance (UXO). TheMPV allows remediation work in treed and rough terrains
where other instruments cannot be deployed; it can work in survey mode and in a static mode for close interrogation of
anomalies. By measuring the three components of the secondary field at five different locations, the MPV provides diverse
time-domain data of high quality. TheMPV is currently being upgraded, streamlined, and enhanced to make it more practical
and serviceable. The new sensor, dubbedMPV-II, has a smaller head and lighter components for better portability. The
original laser positioning system has been replaced with one that uses the transmitter coil as a beacon. The receivers have
been placed in a configuration that permits experimental computation of field gradients. In this work, after introducing the
new sensor, we present the results of several identification/discrimination experiments using data provided by the MPV-II
and digested using a fast and accurate new implementation of the dipole model. The model performs a nonlinear search for
the location of a responding target, at each step carrying out a simultaneous linear least-squares inversion for the principal
polarizabilities at all time gates and for the orientation of the target. We find that the MPV-II can identify standard-issue
UXO, even in cases where there are two targets in its field of view, and can discriminate them from clutter.
Inversion of dynamically repositioned multi-axis electromagnetic data for ordnance characterization
Show abstract
The challenges associated with removing UXO and explosive remnants of war have led to a variety of methods for
detection and discrimination of buried metallic objects using time-domain electromagnetic induction (EMI). Recent
work has shown that parameters recovered from physics-based inversions can discriminate and classify buried ordnance
from non-ordnance. We present results of applying advanced processing to data from a dynamically repositioned multiaxis
EMI instrument. Data are collected using an adaptive sampling process to find the center of the anomaly and collect
minimal data while maintaining model fidelity. An ortho-normalized volume magnetic source (ONVMS) model is used
to resolve various targets at different depths. The ONVMS model is a generalized volume dipole model, with the single
dipole model being a special limiting case. Using the ONVMS model, an object's response to a sensor's primary
magnetic field is modeled mathematically by a set of equivalent magnetic dipoles distributed inside a volume containing
the object. We assess the utility and veracity of the dynamic sampling strategy coupled with the ONVMS model on data
acquired over a set of calibration and simulant targets. Rapid target characterization codes are aggregated into a software
package with particular focus on ease of use for non-expert users.
Electromagnetic Induction II
Live-site UXO classification studies using advanced EMI and statistical models
Show abstract
In this paper we present the inversion and classification performance of the advanced EMI inversion, processing and discrimination
schemes developed by our group when applied to the ESTCP Live-Site UXO Discrimination Study carried out at the former Camp
Butner in North Carolina. The advanced models combine: 1) the joint diagonalization (JD) algorithm to estimate the number of
potential anomalies from the measured data without inversion, 2) the ortho-normalized volume magnetic source (ONVMS) to
represent targets' EMI responses and extract their intrinsic "feature vectors," and 3) the Gaussian mixture algorithm to classify buried
objects as targets of interest or not starting from the extracted discrimination features. The studies are conducted using cued datasets
collected with the next-generation TEMTADS and MetalMapper (MM) sensor systems. For the cued TEMTADS datasets we first
estimate the data quality and the number of targets contributing to each signal using the JD technique. Once we know the number of
targets we proceed to invert the data using a standard non-linear optimization technique in order to determine intrinsic parameters
such as the total ONVMS for each potential target. Finally we classify the targets using a library-matching technique. The
MetalMapper data are all inverted as multi-target scenarios, and the resulting intrinsic parameters are grouped using an unsupervised
Gaussian mixture approach. The potential targets of interest are a 37-mm projectile, an M48 fuze, and a 105-mm projectile. During
the analysis we requested the ground truth for a few selected anomalies to assist in the classification task. Our results were scored
independently by the Institute for Defense Analyses, who revealed that our advanced models produce superb classification when
starting from either TEMTADS or MM cued datasets.
Advanced UXO discrimination: resolving multiple targets and overlapping EMI signals
Show abstract
In this paper we employ advanced electromagnetic induction models to resolve multiple targets with overlapping EMI
signals-i.e. to discriminate objects of interest, such as unexploded ordnance (UXO), from innocuous items. The
models include a) a joint diagonalization (JD) technique that takes data from next-generation EMI sensors and uses the
eigenvalues of the multistatic response matrix to estimate the number of potential targets, and b) the orthonormalized
volume magnetic source (ONVMS) model, a physically complete, fast, and accurate forward model whose
representation of a target's intrinsic EMI response is used to extract classification parameters. In the given approach the
overall EMI inversion and classification problem proceeds as follows: first, the JD is applied to the data and the number
of targets is estimated; once this is known, the ONVMS is combined with an optimization technique to yield the
location and orientation of each buried object, as well as the amplitude of its ONVMS. Finally, a total ONVMS is
calculated for each object and used as a discriminant to distinguish between UXO and non-UXO items and between
different kinds of UXO. We illustrate the applicability of our multi-target analysis technique by using it on several teststand
and live-site datasets collected with the TEMTADS sensor array. We end by demonstrating the superior
performance of the ONVMS by applying it to multi-target blind-test data compiled at the Aberdeen Proving Ground
test-stand facility.
Frequency domain electromagnetic induction sensor data feature extraction and processing for improved landmine detection
Show abstract
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.
EMI sensor positioning using a beacon approach
Show abstract
Discrimination of buried exploded ordnance by inversion of electromagnetic data requires accurate sensor positioning.
There are many contaminated areas were dense forest or significant topographic variation reduces accuracy or precludes
use of standard geo-location methods, such as satellite-based Global Positioning System (GPS) and laser tracking
systems (e.g., Robotic Total Station, RTS), as these rely on line of sight. We propose an alternative positioning system
that is based on a beacon principle. The system was developed to survey with the Man-Portable Vector (MPV) EMI
sensor. The magnetic moment of the MPV transmitter can be detected at a relatively large distance. The primary field is
measured from a portable base station comprised of two vector receivers rigidly attached to either ends of a 1.5 meter
horizontal boom. Control tests showed that relative location and orientation could be recovered with centimeter
positional and one degree angular accuracy within a 3-4-meter range and 60-degree aperture (relative to boom transverse
direction), which is more than sufficient to cover any UXO anomaly. This accuracy level satisfies commonly accepted
positional requirement for discrimination. The beacon positioning system can facilitate classification of munitions in any
man-trafficable area and was successfully deployed at a field demonstration.
Hand-held multi-sensor system design dedicated to mine detection
Show abstract
In this study, a general structure for hand-held multi-sensor mine detection system is proposed. Ideal sensor
configuration for multi-sensor mine detection system, requirements of hardware structures, data transfer issues and
operational restrictions are discussed. The properties of a sample system designed according to the proposed structure
are explained and a new graphical user interface is presented.
A Mélange of Techniques
Synthetic aperture acoustic imaging of canonical targets with a 2-15 kHz linear FM chirp
Show abstract
Synthetic aperture image reconstruction applied to outdoor acoustic recordings is presented. Acoustic imaging
is an alternate method having several military relevant advantages such as being immune to RF jamming, superior spatial
resolution, capable of standoff side and forward-looking scanning, and relatively low cost, weight and size when
compared to 0.5 - 3 GHz ground penetrating radar technologies. Synthetic aperture acoustic imaging is similar to
synthetic aperture radar, but more akin to synthetic aperture sonar technologies owing to the nature of longitudinal or
compressive wave propagation in the surrounding acoustic medium. The system's transceiver is a quasi mono-static
microphone and audio speaker pair mounted on a rail 5meters in length. Received data sampling rate is 80 kHz with a 2-
15 kHz Linear Frequency Modulated (LFM) chirp, with a pulse repetition frequency (PRF) of 10 Hz and an inter-pulse
period (IPP) of 50 milliseconds. Targets are positioned within the acoustic scene at slant range of two to ten meters on
grass, dirt or gravel surfaces, and with and without intervening metallic chain link fencing. Acoustic image
reconstruction results in means for literal interpretation and quantifiable analyses. A rudimentary technique
characterizes acoustic scatter at the ground surfaces. Targets within the acoustic scene are first digitally spotlighted and
further processed, providing frequency and aspect angle dependent signature information.
Detection of unintended electromagnetic emissions from super-regenerative receivers
Jake Hertenstein,
S. Jagannathan
Show abstract
The characteristics of unintended electromagnetic emissions from radio receivers are examined and we propose
a novel method for detecting the presence of these emissions without a priori data or training. The chaotic
properties of the internal oscillators from the radio receivers are modeled and used to detect the device emissions.
A second-order self-similarity model is used to estimate the Hurst parameter as a detection threshold. The
method is compared to a typical threshold method and is shown to be a significant improvement in the accuracy
of detection. Upon detection, the received signal strength can be used to locate the device.
Ground target stimulation using a moving microwave source
David Heberlein,
Bohdan Balko,
Ira Kohlberg,
et al.
Show abstract
This effort is directed toward characterizing the response of a target-placed on or near the ground plane-to the radiation
of a microwave source mounted on a moving vehicle. Because of the multipath problems inherent in measuring the
electric field strength on or near the earth's surface, a theoretical model was constructed so that the effectiveness of the
dynamic approach to actuating specific targets could be characterized. This process is repeated for a variety of microwave
source speeds, vehicle-to-target-aspect ratios, and target positions on or above the ground plane. Given a
horizontally polarized beam driven by a horn antenna, the electric field strengths were calculated at different vehicle-totarget-
aspect ratios and at different heights on and above the ground plane for sand and clay soils. These values were
then modeled in MATLAB to provide target stimulation for varied pulse chains. A model with independent pulse effects
and a model based on the cumulative effect of pulses were evaluated.
DS Sentry: an acquisition ASIC for smart, micro-power sensing applications
Show abstract
Unattended ground monitoring that combines seismic and acoustic information can be a highly valuable tool in
intelligence gathering; however there are several prerequisites for this approach to be viable. The first is high
sensitivity as well as the ability to discriminate real threats from noise and other spurious signals. By combining
ground sensing with acoustic and image monitoring this requirement may be achieved. Moreover, the DS Sentry®provides innate spurious signal rejection by the "active-filtering" technique employed as well as embedding some
basic statistical analysis. Another primary requirement is spatial and temporal coverage. The ideal is
uninterrupted, long-term monitoring of an area. Therefore, sensors should be densely deployed and consume very
little power. Furthermore, sensors must be inexpensive and easily deployed to allow dense placements in critical
areas. The ADVIS DS Sentry®, which is a fully-custom integrated circuit that enables smart, micro-power
monitoring of dynamic signals, is the foundation of the proposed system. The core premise behind this technology
is the use of an ultra-low power front-end for active monitoring of dynamic signals in conjunction with a highresolution,
Σ Δ-based analog-to-digital converter, which utilizes a novel noise rejection technique and is only
employed when a potential threat has been detected. The DS Sentry® can be integrated with seismic accelerometers
and microphones and user-programmed to continuously monitor for signals with specific signatures such as impacts,
footsteps, excavation noise, vehicle-induced ground vibrations, or speech, while consuming only microwatts of
power. This will enable up to several years of continuous monitoring on a single small battery while concurrently
mitigating false threats.
Threat detection in desert environment with passive millimeter-wave sensor
Show abstract
A new technique for improvised explosive device (IED) creation uses an explosive device buried in foam and covered in
a layer of dirt. These devices are difficult to detect visually, however, their material characteristics make them detectable
by passive millimeter-wave (pmmW) sensors. Results are presented from a test using a mock IED and an outdoor set-up
consisting of two mock IEDs on a dirt background. The results show that the mock IEDs produces a millimeter-wave
signature which is distinguishable from the background surrounding the mock IEDs. Simulations based on the measured
data are presented and a design for a future vehicle mounted sensor is shown.
Laser neutralization of buried munitions
Show abstract
This report describes the results of the first phase of a planned two-phase program to develop laser technology for rapid neutralization of buried munitions from a safe standoff distance. The primary objective of this first phase is to demonstrate, via laboratory experiments, the capabilities of a breadboard laser system to "drill" through a minimum depth of 15 cm of earthen materials to defeat a buried mine at a standoff distance greater than 20 m. In the initial phase covered by this report, results of short range laboratory testing by 3 contractors are reported. The planned second phase of this program will consist of procuring a more capable 10 kW SM laser and performing field-testing at longer standoff ranges.
Sensing and Detecting in the Marine Environment I
Automation for underwater mine recognition: current trends and future strategy
Show abstract
The purpose of this paper is to define the vision and future strategy for advancing the use of automation in underwater
mine recognition. The technical portion of this strategy is founded on the principle of adapting the automation in situ
based on a highly variable environment / context and the occasional availability of the human operator. To frame this
strategy, a survey of past and current algorithm development for underwater mine recognition is presented and includes a
detailed description on adaptive algorithms. This discussion is motivated by illustrating the extreme variability in the
underwater environment and that performance estimation techniques are now emerging that are capable of quantifying
these variations in situ. It is the in situ linkage of performance estimation with adaptive recognition that forms one of the
key technological enablers of this future strategy. The non-technical portion of this strategy is centered on enabling an
effective human-machine team. Enabling this teaming relationship involves both gaining trust and establishing an
overall support system that is amenable to such human-machine interactions. Aspects of trust include both individual
trust and institutional trust, and a path for gaining both is discussed. Overall aspects of the support system are
highlighted and include standards for data and interoperability, network-centric software architectures, and issues in
proliferating knowledge that is learned in situ by multiple distributed algorithms. This paper concludes with an
articulation of several important and timely research questions concerning automation for underwater mine recognition.
Seabed segmentation in synthetic aperture sonar images
Show abstract
A synthetic aperture sonar (SAS) image segmentation algorithm using features from a parameterized intensity
image autocorrelation function (ACF) is presented. A modification over previous parameterized ACF models that
better characterizes periodic or rippled seabed textures is presented and discussed. An unsupervised multiclass
k-means segmentation algorithm is proposed and tested against a set of labeled SAS images. Segmentation
results using the various models are compared against sand, rock, and rippled seabed environments.
Optimal frame pursuit for pattern classification
Jason C. Isaacs
Show abstract
Frame methods, basis expansion methods, or kernel methods provide a higher-dimensional representation of a given
dataset within a feature space for discrimination applications. Frame pursuit addresses the problem of searching for
optimal frames to improve classification for pattern recognition applications. In this paper, the results of two stochastic
optimization techniques applied to the optimal frame problem are presented. The cost function is a k-nearest-neighbor
function. These techniques are tested here over six datasets. Empirical results demonstrate the utility of frame
transformations for improving performance results in pattern recognition applications.
Statistical analysis and classification of acoustic color functions
Show abstract
In this paper we present a method for clustering and classification of acoustic color data based on statistical
analysis of functions using square-root velocity functions (SVRF). The convenience of the SVRF is that it transforms
the Fisher-Rao metric into the standard L2 metric. As a result, a formal distance can be calculated using
geodesic paths. Moreover, this method allows optimal deformations between acoustic color data to be computed
for any two targets allowing for robustness to measurement error. Using the SVRF formulation statistical models
can then be constructed using principal component analysis to model the functional variation of acoustic color
data. Empirical results demonstrate the utility of functional data analysis for improving performance results in
pattern recognition using acoustic color data.
Seabed change detection in challenging environments
Cameron A. Matthews,
Daniel D. Sternlicht
Show abstract
Automatic Change Detection (ACD) compares new and stored terrain images for alerting to changes occurring
over time. These techniques, long used in airborne radar, are just beginning to be applied to sidescan sonar. Under the right
conditions ACD by image correlation-comparing multi-temporal image data at the pixel or parcel level-can be used to detect
new objects on the seafloor. Synthetic aperture sonars (SAS)-coherent sensors that produce fine-scale, range-independent
resolution seafloor images-are well suited for this approach; however, dynamic seabed environments can introduce "clutter"
to the process. This paper explores an ACD method that uses salience mapping in a global-to-local analysis architecture. In
this method, termed Temporally Invariant Saliency (TIS), variance ratios of median-filtered repeat-pass images are used to
detect new objects, while deemphasizing modest environmental or radiometric-induced changes in the background. Successful
tests with repeat-pass data from two SAS systems mounted on autonomous undersea vehicles (AUV) demonstrate the
feasibility of the technique.
Sensing and Detecting in the Marine Environment II
Laplace-Beltrami eigenfunctions for 3D shape matching
Jason C. Isaacs
Show abstract
Assuming that a 2D surface is a representation of a manifold embedded in 3-space then metrics of the eigenfunctions of
the diffusion maps of that manifold represent the shape of that manifold with invariance to rotation, scale, and
translation. Diffusion maps is said to preserve the local proximity between data points by constructing a representation
for the underlying manifold by an approximation of the Laplace-Beltrami operator acting on the graph of this
surface. This work examines 3D shape clustering problems using metrics of the projections onto the natural and nodal
sets of the Laplace-Betrami eigenfunctions for shape analysis of closed surfaces. Results demonstrate that the metrics
allow for good class separation over multiple targets with noise.
Deformable Bayesian networks for data clustering and fusion
Show abstract
In this work, we propose DEformable BAyesian Networks (DEBAN), a probabilistic graphical model framework
where model selection and statistical inference can be viewed as two key ingredients in the same iterative
process. While this concept has shown successful results in computer vision community,1-4 our proposed approach
generalizes the concept such that it is applicable to any data type. Our goal is to infer the optimal structure/model
to fit the given observations. The optimal structure conveys an automatic way to find not only the number of
clusters in the data set, but also the multiscale graph structure illustrating the dependence relationship among
the variables in the network. Finally, the marginal posterior distribution at each root node is regarded as the
fused information of its corresponding observations, and the most probable state can be found from the maximum
a posteriori (MAP) solution with the uncertainty of the estimate in the form of a probability distribution which
is desired for a variety of applications.
Bayesian surprise metric for outlier detection in on-line learning
Show abstract
Our previous work developed an online learning Bayesian framework (dynamic tree) for data organization and
clustering. To continuously adapt the system during operation, we concurrently seek to perform outlier detection
to prevent them from incorrectly modifying the system. We propose a new Bayesian surprise metric to differentiate
outliers from the training data and thus help to selectively adapt the model parameters. The metric is
calculated based on the difference between the prior and the posterior distributions on the model when a new
sample is introduced. A good training datum would sufficiently but not excessively change the model; consequently,
the difference between the prior and the posterior distributions would be reasonable to the amount of
new information present on the datum. However, an outlier carries an element of surprise that would significantly
change the model. In such a case, the posterior distribution would greatly differ from the prior resulting in a large
value for the surprise metric. We categorize this datum as an outlier and other means (e.g. human operator) will
have to be used to handle such cases. The surprise metric is calculated based on the model distribution, and as
such, it adapts with the model. The surprise factor is dependent on the state of the system. This speeds up the
learning process by considering only the relevant new data. Both the model parameters and even the structure
of the dynamic tree can be updated under this approach.
Low-noise magnetic sensing for marine munitions characterization
Show abstract
Because the recovery of underwater munitions is many times more expensive than recovering the same items on dry
land, there is a continuing need to advance marine geophysical characterization methods. To efficiently and reliably
conduct surveying in marine environments, low-noise geophysical sensors are being configured to operate close to the
sea bottom. We describe systems that are deployed from surface vessels via rigid or flexible tow cables or mounted
directly to submersible platforms such as unmanned underwater vehicles. Development and testing of a towed
configuration has led to a 4 meter wide hydrodynamically stable tow wing with an instrumented top-side assembly
mounted on the stern of a surface survey vessel. An integrated positioning system combined with an instrumented cable
management system, vessel and wing attitude and wing depth measurements to provide sub-meter positional accuracy in
up to 25 meter water depths and within 1 to 2 meters of the seafloor. We present the results of data collected during an
instrument validation survey over a series of targets emplaced at measured locations. Performance of the system was
validated through analyses of data collected at varying speeds, headings, and height above the seafloor. Implementation
of the system during live-site operations has demonstrated its capability to survey hundreds of acres of marine or
lacustrine environment. Unique deployment concepts that utilize new miniaturized and very low noise sensors show
promise for expanding the applicability of magnetic sensing at marine sites.
Active source electromagnetic methods for marine munitions
Show abstract
The detection of munitions targets obscured in coastal and marine settings has motivated the need for advanced
geophysical technologies suited for underwater deployment. Building on conventional marine electromagnetic theory
and based on the use of existing electric and magnetic field sensing designs, we analyze the electromagnetic fields
emitted from excited targets in the frequency range between 1 kHz and 1 MHz. We present evidence that employing
electromagnetic modes that are higher in frequency relative to those typically used in ground-based sensing yields
greater range and sensitivity for underwater surveys. We develop potential design strategies for implementing both
magnetic (B) and electric (E) field sources and sensors in the marine environment, and determine optimal arrangements
for a potential combined E- and B-field sensing system. The implementation of both 1D analytical and 3D numerical
simulations yields the primary and secondary field distributions in representative underwater settings for various sourcereceiver
arrangements. We study the electromagnetic field distributions from both electric (voltage-fed dipole) and
magnetic field (encased and submerged induction coil) active sources. Application of these concepts provide unique and
useful information about targets from the addition of electric field sensing alone as well as through the combination of
electric and magnetic field sensing.
Investigating Tx coils and magnetic field Rx sensor configurations for underwater geo-location
Show abstract
In this work, new configurations of magnetic field transmitter coils (Tx) and receiver sensors (Rx) are studied for underwater (UW)
geo-locations. The geo-location system, based on low frequency magnetic fields, uses measured vector magnetic fields at a given set
of points in space. It contains an active pulsed direct current transmitter, tri-axial field receivers, and a global positioning system unit
(GPS). The GPS is coupled with the EMI system and provides continuous geo-referencing of the UW system's position. UW geolocations
are estimated using a) closed form solution, that uses the total vector magnetic field tensor's gradient, and b) nonlinear
optimization technique based of the differential evolution (DE) algorithm. In this work we first investigated the advantages and
disadvantages of the proposed UW low frequency magnetic field geo-location system. Namely, we present systematic studies on: a)
magnetic field transmitter configurations to determine the best compromise between size, shape and practical implementation to
achieve maximum transmitter range in the UW environment, b) the placements of tri-axial receiver sensors with respect to the Tx to
accurately estimate the UW geo-location from the measured magnetic fields; c) different sources of noise (such as the air-water
interface, coupling between targets' EMI responses and the geo-location system's signals, water conductivity), to estimate how these
noises influence the system's performance and localization precision. Finally, we assessed the capabilities of the closed-form solution
and the DE technique to predict the location of an underwater interrogation system by comparing their corresponding estimated results
to the true value. We found that for realistic water conductivities, the frequency should be less than 100 Hz. We showed that when the
primary magnetic field is contaminated with random noises due to the presence of underwater metallic targets, water
conductivity/frequency changes, and finite size of the transmitter, the performance of the full vector magnetic field tensor gradient
approach degrades significantly compared to that of the DE method. In addition, the number of receivers required by the vector
magnetic field tensor gradient technique and its sensitivity with respect to the sensor separation prevented us from further considering
this technique for UW geo-location, leaving the non-linear approach, that uses only three vector Rx, as our technique of choice for
tracking the location of underwater interrogation sensors with centimeter-level accuracy.
Soils and Soil Effects
Coaxial line measurement and analysis of electromagnetic properties of soils for sensor applications
Show abstract
We report complex permittivity, conductivity, magnetic susceptibility, and attenuation for soils collected from a typical
site in a current theater of operations. Our experimental setup consists of three network analyzers along with custombuilt
sample holders and data reduction and analysis software. The sample holder has the advantage of large sample
volume and a resulting higher signal to noise ratio. This system was developed to determine the electrical properties of
soils over a wide frequency range from 100 Hz to 8 GHz. The lower frequencies are applicable to capacitive sensors for
small shallow targets, while the higher frequencies are applicable to ground-penetrating radar (GPR) from 50 MHz to 2
GHz and beyond. S-parameter data is collected and reduced using a method, initially developed by Nicolson and Ross
(1970)1, for the determination of dielectric permittivity, magnetic permeability, and loss tangent from measured Sparameter
data. Experimental results are compared with site geology and mineralogy. Applications include detection of
tunnels, land mines, unexploded ordinance (UXO), concrete reinforcements, and other shallow compact targets.
Performance of demining sensors and soil properties
Show abstract
Metal detector has commonly been used for landmine detection and ground-penetrating radar (GPR) is about to be
deployed as dual sensor that is in combination with metal detector. Since both devices employ electromagnetic
techniques, they are influenced by magnetic and dielectric properties of soil. To observe the influence, various soil
properties as well as their spatial distributions were measured in four types of soil where a field test of metal detectors
and GPRs took place. By analyzing soil properties these four types of soil were graded based on the estimated amount of
influence on the detection techniques. The classification was compared to the detection performance of devices obtained
from the blind test and a clear correlation between the difficulty of soil and the performance was observed; the detection
and identification performance were degraded in soils that were classified as problematic. Therefore, it was demonstrated
that the performance of metal detector and GPR for landmine detection can qualitatively be assessed by geophysical
analyses.
Effects of different soil types on strip-map GPR SAR images
Show abstract
In this study, we present generation of Strip-map Synthetic Aperture Radar (SAR) images using impulse GPR system,
and investigate effects of different soil types on SAR images. The SAR images of buried objects have been interpreted
via 2D inverse Fourier transformation. GPR buried target data have been collected from three soil pools having different
dielectric constants and B-scan images have been reconstructed from the received data using mean A-scan signal
subtraction method. In order to reconstruct SAR images, the time domain data collected from multiple observation
points have been transformed to 2D spectral domain. Non-uniform data have been interpolated over spatial Cartesian
grid by using uniform interval. Thus, the SAR images have been reconstructed via 2D inverse FFT of interpolated data
on ky-kz plane. When examined mathematical background of SAR algorithm, the values of different dielectric constants
change the wave number of k. This can lead to deterioration of the SAR imagery. In this study, we investigate the Effect
of the dielectric constant of different soils has been examined on SAR images. Finally, resolution difference between
background removed B-Scan data and SAR images is considered.
High-resolution soil moisture mapping in Afghanistan
Show abstract
Soil moisture conditions have an impact upon virtually all aspects of Army activities and are increasingly affecting its
systems and operations. Soil moisture conditions affect operational mobility, detection of landmines and unexploded
ordinance, natural material penetration/excavation, military engineering activities, blowing dust and sand, watershed
responses, and flooding. This study further explores a method for high-resolution (2.7 m) soil moisture mapping using
remote satellite optical imagery that is readily available from Landsat and QuickBird. The soil moisture estimations are
needed for the evaluation of IED sensors using the Countermine Simulation Testbed in regions where access is difficult
or impossible. The method has been tested in Helmand Province, Afghanistan, using a Landsat7 image and a QuickBird
image of April 23 and 24, 2009, respectively. In previous work it was found that Landsat soil moisture can be predicted
from the visual and near infra-red Landsat bands1-4. Since QuickBird bands 1-4 are almost identical to Landsat bands 1-
4, a Landsat soil moisture map can be downscaled using QuickBird bands 1-4. However, using this global approach for
downscaling from Landsat to QuickBird scale yielded a small number of pixels with erroneous soil moisture values.
Therefore, the objective of this study is to examine how the quality of the downscaled soil moisture maps can be
improved by using a data stratification approach for the development of downscaling regression equations for each
landscape class. It was found that stratification results in a reliable downscaled soil moisture map with a spatial
resolution of 2.7 m.
Detection of Bulk Explosives Threats I
Principles and status of neutron-based inspection technologies
Tsahi Gozani
Show abstract
Nuclear based explosive inspection techniques can detect a wide range of substances of importance for a wide range
of objectives. For national and international security it is mainly the detection of nuclear materials, explosives and
narcotic threats. For Customs Services it is also cargo characterization for shipment control and customs duties. For
the military and other law enforcement agencies it could be the detection and/or validation of the presence of
explosive mines, improvised explosive devices (IED) and unexploded ordnances (UXO).
The inspection is generally based on the nuclear interactions of the neutrons (or high energy photons) with the
various nuclides present and the detection of resultant characteristic emissions. These can be discrete gamma lines
resulting from the thermal neutron capture process (n,γ) or inelastic neutron scattering (n,n'γ) occurring with fast
neutrons. The two types of reactions are generally complementary. The capture process provides energetic and
highly penetrating gamma rays in most inorganic substances and in hydrogen, while fast neutron inelastic scattering
provides relatively strong gamma-ray signatures in light elements such as carbon and oxygen. In some specific
important cases unique signatures are provided by the neutron capture process in light elements such as nitrogen,
where unusually high-energy gamma ray is produced. This forms the basis for key explosive detection techniques.
In some cases the elastically scattered source (of mono-energetic) neutrons may provide information on the atomic
weight of the scattering elements.
The detection of nuclear materials, both fissionable (e.g., 238U) and fissile (e.g., 235U), are generally based on the
fissions induced by the probing neutrons (or photons) and detecting one or more of the unique signatures of the
fission process. These include prompt and delayed neutrons and gamma rays. These signatures are not discrete in
energy (typically they are continua) but temporally and energetically significantly different from the background,
thus making them readily distinguishable.
The penetrability of neutrons as probes and signatures as well as the gamma ray signatures make neutron
interrogation applicable to the inspection of large conveyances such as cars, trucks, marine containers and also
smaller objects like explosive mines concealed in the ground.
The application of nuclear interrogation techniques greatly depends on operational requirements. For example
explosive mines and IED detection clearly require one-sided inspection, which excludes transmission based
inspection (e.g., transmission radiography) and greatly limits others.
The technologies developed over the last decades are now being implemented with good results. Further advances
have been made over the last several years that increase the sensitivity, applicability and robustness of these systems.
The principle, applications and status of neutron-based inspection techniques will be reviewed.
Portable and autonomous X-ray equipment for in-situ threat materials identification by effective atomic number high-accuracy measurement
Show abstract
A novel portable and autonomous X-ray dual-energy Radioscopy equipment, developed for bomb squad interventions
and NDT applications and capable of in-situ digital radiography imaging with measurement of the effective Atomic
number of materials (Zeff), is presented. The system consists of a 2D dual-energy X-ray detector based on a rapidly
translated linear array, a portable X-ray source and dedicated software running on a laptop or tablet PC. By measurement
of the collected x-ray intensities at two different energy spectra, the system can directly compute the material Zeff value
for various organic materials contained in the scanned object and then identify them from a database list. The entire
system calibration has been obtained using explosive simulants with known Zeff values, the measurement error of Zeffbeing around ±3.5 % with respect to the reference values. The excellent image resolution and the ability of the automated
threat identification algorithm are presented for experiments with a briefcase and a hand-held baggage having various
domestic objects and an explosive simulant inside.
Defence Research and Development Canada: Suffield research on nuclear methods for detection of buried bulk explosives
Show abstract
Defence R&D Canada - Suffield has conducted research and development on nuclear methods for detection of
bulk explosives since 1994. Initial efforts were directed at confirmation of the presence of bulk explosives in
land mines and improvised explosive devices (IEDs). In close collaboration with a few key Canadian companies,
methods suitable for vehicle-mounted or fixed position applications and those suitable for person- or small robotportable
roles have been studied. Vehicle-mounted systems mainly employ detection of characteristic radiation,
whereas person-portable systems use imaging of back scattered radiation intensity distributions. Two key design
tenets have been reduction of personnel shielding by the use of teleoperation and custom design of sensors to
address the particular problem, rather than adapting an existing sensor to a problem. This is shown in a number
of recent research examples.
Among vehicle-mounted systems, recent research to improve the thermal neutron analysis (TNA) sensors,
which were put into service with the Canadian Forces in 2002, are discussed. Research on fast neutron analysis
(FNA) and associated particle imaging (API), which can augment or replace TNA, depending on the application,
are described. Monoenergetic gamma ray induced photoneutron spectroscopy is a novel method which has a
number of potential advantages and disadvantages over TNA and FNA. Sources, detectors and geometries have
been identified and modelling studies have suggested feasibility. Among person-portable systems, research on
neutron backscatter imaging and X-ray coded aperture backscatter imaging are discussed.
Detection of Bulk Explosives Threats II
Nuclear quadrupole resonance detection of explosives: an overview
Joel B. Miller
Show abstract
Nuclear Quadrupole Resonance (NQR) is a spectroscopic technique closely related to Nuclear Magnetic
Resonance (NMR) and Magnetic Resonance Imaging (MRI). These techniques, and NQR in particular, induce signals
from the material being interrogated that are very specific to the chemical and physical structure of the material, but are
relatively insensitive to the physical form of the material. NQR explosives detection exploits this specificity to detect
explosive materials, in contrast to other well known techniques that are designed to detect explosive devices. The past
two decades have seen a large research and development effort in NQR explosives detection in the United States aimed
at transportation security and military applications. Here, I will briefly describe the physical basis for NQR before
discussing NQR developments over the past decade, with particular emphasis on landmine detection and the use of NQR
in combating IED's. Potential future directions for NQR research and development are discussed.
Observations on military exploitation of explosives detection technologies
Show abstract
Accurate and timely detection of explosives, energetic materials, and their associated compounds would provide
valuable information to military commanders in a wide range of military operations: protection of fast moving
convoys from mobile or static IED threats; more deliberate countermine and counter-IED operations during
route or area clearance; and static roles such as hasty or deliberate checkpoints, critical infrastructure protection
and support to public security.
The detection of hidden explosive hazards is an extremely challenging problem, as evidenced by the fact that
related research has been ongoing in many countries for at least seven decades and no general purpose solution
has yet been found. Technologies investigated have spanned all major scientific fields, with emphasis on the
physical sciences, life sciences, engineering, robotics, computer technology and mathematics.
This paper will present a limited, operationally-focused overview of the current status of detection technologies.
Emphasis will be on those technologies that directly detect the explosive hazard, as opposed to those that
detect secondary properties of the threat, such as the casing, associated wires or electronics. Technologies that
detect explosives include those based on nuclear radiation and terahertz radiation, as well as trace and biological
detection techniques. Current research areas of the authors will be used to illustrate the practical applications.
Explosives (and other threats) detection using pulsed neutron interrogation and optimized detectors
Dan A. Strellis,
Mashal Elsalim,
Tsahi Gozani
Show abstract
We have previously reported results from a human-portable system using neutron interrogation to detect contraband and
explosives. We summarized our methodology for distinguishing threat materials such as narcotics, C4, and mustard gas
in the myriad of backgrounds present in the maritime environment. We are expanding our mission for the Domestic
Nuclear Detection Office (DNDO) to detect Special Nuclear Material (SNM) through the detection of multiple fission
signatures without compromising the conventional threat detection performance.
This paper covers our initial investigations into using neutrons from compact pulsed neutron generators via the
d(D,n)3He or d(T,n)α reactions with energies of ~2.5 and 14 MeV, respectively, for explosives (and other threats)
detection along with a variety of gamma-ray detectors. Fast neutrons and thermal neutrons (after successive collisions)
can stimulate the emission of various threat detection signatures. For explosives detection, element-specific gamma-ray
signatures via the (n,n'γ) inelastic scattering reaction and the (n,'γ) thermal capture reaction are detected. For SNM,
delayed gamma-rays following fission can be measured with the same detector. Our initial trade-off investigations of
several gamma-ray detectors types (NaI, CsI, LaBr3, HPGe) for measuring gamma-ray signatures in a pulsed neutron
environment for potential application in a human-portable active interrogation system are covered in this paper.
A non-imaging polarized terahertz passive system for detecting and identifying concealed explosives
Show abstract
Existing terahertz THz systems for detecting concealed explosives are not capable of identifying explosive type
which leads to higher false alarm rates. Moreover, some of those systems are imaging systems that invade personal
privacy, and require more processing and computational resources. Other systems have no polarization preference
which makes them incapable of capturing the geometric features of an explosive.
In this study a non-imaging polarized THz passive system for detecting and identifying concealed explosives
overcoming the forgoing shortcomings is developed. The system employs a polarized passive THz sensor in
acquiring emitted data from a scene that may have concealed explosives. The acquired data are decomposed into
their natural resonance frequencies, and the number of those frequencies is used as criteria in detecting the explosive
presence. If the presence of an explosive is confirmed, a set of physically based retrieval algorithms is used in
extracting the explosive dielectric constant/refractive index value from natural resonance frequencies and amplitudes
of associated signals. Comparing the refractive index value against a database of refractive indexes of known
explosives identifies the explosive type.
As an application, a system having a dual polarized radiometer operating within the frequency band of 0.62- 0.82
THz is presented and used in detecting and identifying person borne C-4 explosive concealed under a cotton
garment. The system showed higher efficiencies in detecting and identifying the explosive.
Standoff Detection of Explosives
Detection and identification of explosives hidden under barrier materials: what are the THz-technology challenges?
Show abstract
We describe experiments where different explosives were hidden under common barrier materials, and THz radiation
was used to detect and identify these explosives. Our THz system, a time-domain spectroscopy (TDS) system, is based
on a femtosecond laser whose radiation is converted into THz radiation by a low-temperature grown GaAs
photoconductive switch. A similar switch detects the reflected signal. The advantage of using a TDS system is that
pulses reflected from the barrier and the actual explosive, arrive at different instances at the detector. This simplifies the
separation of the barrier signature from the explosive signature, compared to a frequency domain system. However,
partial temporal overlap between the two pulses makes it challenging to completely separate the spectral characteristics
of the explosive from the characteristics of the barrier. Also, in addition to attenuating the THz-pulses, transmission
through barrier materials may add spectral features to the reflected signal, hampering recognition of the explosive. On
top of that, the explosive may have a rough surface, which reduces the strength of the reflected signal. In this
contribution we shall address these issues and discuss strategies that may be used to face these challenges.
Stand-off detection of explosive particles by imaging Raman spectroscopy
Show abstract
A multispectral imaging technique has been developed to detect and identify explosive particles, e.g. from a fingerprint,
at stand-off distances using Raman spectroscopy.
When handling IED's as well as other explosive devices, residues can easily be transferred via fingerprints onto other
surfaces e.g. car handles, gear sticks and suite cases. By imaging the surface using multispectral imaging Raman
technique the explosive particles can be identified and displayed using color-coding.
The technique has been demonstrated by detecting fingerprints containing significant amounts of 2,4-dinitrotoulene
(DNT), 2,4,6-trinitrotoulene (TNT) and ammonium nitrate at a distance of 12 m in less than 90 seconds
(22 images × 4 seconds)1. For each measurement, a sequence of images, one image for each wave number, is recorded.
The spectral data from each pixel is compared with reference spectra of the substances to be detected. The pixels are
marked with different colors corresponding to the detected substances in the fingerprint.
The system has now been further developed to become less complex and thereby less sensitive to the environment such
as temperature fluctuations. The optical resolution has been improved to less than 70 μm measured at 546 nm
wavelength. The total detection time is ranging from less then one minute to around five minutes depending on the size
of the particles and how confident the identification should be. The results indicate a great potential for multi-spectral
imaging Raman spectroscopy as a stand-off technique for detection of single explosive particles.
Picosecond laser pulses improves sensitivity in standoff explosive detection
Madeleine Åkeson,
Markus Nordberg,
Anneli Ehlerding,
et al.
Show abstract
Portendo has in collaboration with FOI, the Swedish Defence Research Agency, developed a world-leading
technique of trace detection of explosives at standoff distance using Raman spectroscopy. The technology is
further developed in order to enhance the sensitivity of the method and to be able to extend the field of
applications. Raman scattering is a well-established technique able to detect substances down to single
micrograms at standoff distances, however, one of the obstacles limiting the detection possibilities is interfering
fluorescence, originating either from the substance itself or from the surrounding material. One main challenge
for this technology is thus to either omit the excitation of the fluorescent process altogether or to be able to
separate the two processes and only detect the Raman signal.
Due to the difference in the temporal behavior of the two processes - Raman scattering occurs in the order of
femtoseconds while fluorescence typically has a lifetime in the order of nanoseconds - one way to theoretically
separate them is to limit the measurement to as short time as possible, cutting off most of the emitted
fluorescence. The improvement depends on how much of the fluorescence can be omitted without decreasing
the Raman signal. Experimentally, we have verified this improvement in signal to noise ratio when using a laser
with picosecond pulses instead of nanosecond pulses, which has resulted in an improvement in SNR of up to 7
times for bulk ANFO. These results verify the predicted signal enhancement and suggest higher sensitivity for
standoff detection in future systems.
Advances in Ground Penetrating Radar Subsurface Object Detection
Random GPR antennae height variations and mine detection performance
Show abstract
A database of vehicle mounted GPR mine data acquired over rough road conditions is analyzed to determine the
sensitivity to random and rapid variations in antenna heights due to naturally occurring antenna bounce. Results are
also described for data acquired during tests designed to induce specific bounce profiles during data collections.
Significant increases are observed in false alarms, and significant decreases are observed in mine detection
probabilities at rougher test lane locations. Perhaps the most significant antenna height factor was discovered to be
rapid and large GPR amplitude changes due to the variations in antenna height and thus propagation loss. Finally,
normalization techniques developed by AARD are examined to correct for these GPR antenna bounce factors and are
shown to be effective in most cases, particularly for some of the more severe induced bounce scenarios.
Detection of explosive hazards using spectrum features from forward-looking ground penetrating radar imagery
Show abstract
Buried explosives have proven to be a challenging problem for which ground penetrating radar (GPR) has shown to be
effective. This paper discusses an explosive hazard detection algorithm for forward looking GPR (FLGPR). The
proposed algorithm uses the fast Fourier transform (FFT) to obtain spectral features of anomalies in the FLGPR
imagery. Results show that the spectral characteristics of explosive hazards differ from that of background clutter and are
useful for rejecting false alarms (FAs). A genetic algorithm (GA) is developed in order to select a subset of spectral
features to produce a more generalized classifier. Furthermore, a GA-based K-Nearest Neighbor probability density
estimator is employed in which targets and false alarms are used as training data to produce a two-class classifier. The
experimental results of this paper use data collected by the US Army and show the effectiveness of spectrum based
features in the detection of explosive hazards.
Narrow-band processing and fusion approach for explosive hazard detection in FLGPR
Show abstract
This paper proposes an effective anomaly detection algorithm for a forward-looking ground-penetrating radar(FLGPR). One challenge for threat detection using FLGPR is its high dynamic range in response to different kinds of
targets and clutter objects. The application of a fixed threshold for detection in a full-band radar image often yields a
large number of false alarms. We propose a method that uses both narrow-band and full-band radar processing, coupled
with a classifier that uses complex-valued Gabor filter responses as the features. We then fuse the narrow-band and fullband
images into a composite confidence map and detect local maxima in this map to produce candidate alarm
locations. Full-band radar images provide a high degree of image resolution, while narrow-band images provide a means
to detect targets which have a unique narrow-band signature. Experimental results for our improved detection
techniques are demonstrated on data sets collected at a US Army test site.
Layer segmentation of GPR images using relaxation labeling for landmine detection
Show abstract
A method for segmenting deformable shapes of soil and ground layers in successive GPR image frames is described in
this paper. First, pre-processing operators are applied to enhance the quality of each image frame. Second, local
histogram features are used to initialize membership probabilities of each pixel in the current image frame. Then, a
segmentation algorithm based on relaxation labeling is applied to perform image segmentation. This algorithm uses
information from previous and current image frames to perform layer identification by formulating the segmentation task
as a probabilistic relaxation labeling process in which the current frame image is used for initializing pixel membership
probabilities estimated from gray-level histograms. The previous frame image is used for estimating the compatibility
values to be utilized for segmenting the current frame image using mutual information among neighboring pixels. By
iteratively refining the membership probabilities of each pixel in the current image frame in parallel, an enhanced
segmentation is produced according to the refined probabilities. A distinguishing characteristic of this process is the
ability to incorporate both temporal contexts (down-track history information encoded as compatibilities) and spatial
contexts (current-scan pixel neighborhood information encoded as probabilities), concurrently. The segmented image is
post-processed by further filtering operations and checking for highly unlikely decisions to produce the final
segmentation.
Tracking Rough Ground in Ground-Penetrating Radar Data
The Viterbi algorithm as an approach for incorporating spatial information into air/ground interface inference in GPR data
Show abstract
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.
DynaMax+ ground-tracking algorithm
Show abstract
In this paper, we propose a new method for performing ground-tracking using ground-penetrating radar (GPR).
Ground-tracking involves identifying the air-ground interface, which is usually the dominant feature in a radar
image but frequently is obscured or mimicked by other nearby elements. It is an important problem in landmine
detection using vehicle-mounted systems because antenna motion, caused by bumpy ground, can introduce
distortions in downtrack radar images, which ground-tracking makes it possible to correct. Because landmine
detection is performed in real-time, any algorithm for ground-tracking must be able to run quickly, prior to other,
more computationally expensive algorithms for detection. In this investigation, we first describe an efficient
algorithm, based on dynamic programming, that can be used in real-time for tracking the ground. We then
demonstrate its accuracy through a quantitative comparison with other proposed ground-tracking methods, and
a qualitative comparison showing that its ground-tracking is consistent with human observations in challenging
terrain.
Support vector data description for detecting the air-ground interface in ground penetrating radar signals
Show abstract
In using GPR images for landmine detection it is often useful to identify the air-ground interface in the GRP
signal for alignment purposes. A common simple technique for doing this is to assume that the highest return
in an A-scan is from the reflection due to the ground and to use that as the location of the interface. However
there are many situations, such as the presence of nose clutter or shallow sub-surface objects, that can cause the
global maximum estimate to be incorrect. A Support Vector Data Description (SVDD) is a one-class classifier
related to the SVM which encloses the class in a hyper-sphere as opposed to using a hyper-plane as a decision
boundary. We apply SVDD to the problem of detection of the air-ground interface by treating each sample in
an A-scan, with some number of leading and trailing samples, as a feature vector. Training is done using a set
of feature vectors based on known interfaces and detection is done by creating feature vectors from each of the
samples in an A-scan, applying the trained SVDD to them and selecting the one with the least distance from
the center of the hyper-sphere. We compare this approach with the global maximum approach, examining both
the performance on human truthed data and how each method affects false alarm and true positive rates when
used as the alignment method in mine detection algorithms.
Comparison of algorithms for finding the air-ground interface in ground penetrating radar signals
Show abstract
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.
Signal Processing Ground-Penetrating Radar Data I
Observations on syntactic landmine detection using impulse ground-penetrating radar
Show abstract
We discuss some results and observations on applying syntactic pattern recognition (SPR) methodology for
landmine detection using impulse ground-penetrating radar (GPR). In the SPR approach, the GPR A-scans
are first converted into binary-valued strings by inverse filtering, followed by concavity detection to identify
the peaks and valleys representing the locations of impedance discontinuities in the return signal. During
the training phase, the characteristic binary strings for a particular landmine are found by looking at all
the exemplars of that mine and selecting the collection of strings that yield the best detection results on
these exemplars. These characteristic strings can be detected very efficiently using finite state machines
(FSMs). Finally, the FSM detections are clustered to assign confidence to each detection, and discard sparse
detections.
Provided that the impulse GPR provides enough resolution in range, the SPR method can be a robust
and high-speed solution for landmine detection and classification, because it aims to exploit the impedance
discontinuity profile of the target, which is a description of the internal material structure of the target and
little affected by external clutter. To evaluate the proposed methodology, the SPR scheme is applied to a set
of impulse GPR data taken at a government test site. We suggest that coherent frequency-agile radar may
be a better option for the SPR approach, since it addresses some of the drawbacks of a non-coherent impulse
GPR caused by internally non-coherent within-channel signals which necessitate non-coherent integration
and its attendant longer integration times, and non-coherent adjacent channels which severely limit the
ability to do spatial, or at a minimum, cross-range processing if the GPR is in a linear array antenna.
Characterization of binary string statistics for syntactic landmine detection
Show abstract
Syntactic landmine detection has been proposed to detect and classify non-metallic landmines using ground
penetrating radar (GPR). In this approach, the GPR return is processed to extract characteristic binary
strings for landmine and clutter discrimination. In our previous work, we discussed the preprocessing methodology
by which the amplitude information of the GPR A-scan signal can be effectively converted into binary
strings, which identify the impedance discontinuities in the signal.
In this work, we study the statistical properties of the binary string space. In particular, we develop a
Markov chain model to characterize the observed bit sequence of the binary strings. The state is defined
as the number of consecutive zeros between two ones in the binarized A-scans. Since the strings are highly
sparse (the number of zeros is much greater than the number of ones), defining the state this way leads to
fewer number of states compared to the case where each bit is defined as a state. The number of total states
is further reduced by quantizing the number of consecutive zeros. In order to identify the correct order of
the Markov model, the mean square difference (MSD) between the transition matrices of mine strings and
non-mine strings is calculated up to order four using training data. The results show that order one or two
maximizes this MSD.
The specification of the transition probabilities of the chain can be used to compute the likelihood of
any given string. Such a model can be used to identify characteristic landmine strings during the training
phase. These developments on modeling and characterizing the string statistics can potentially be part of a
real-time landmine detection algorithm that identifies landmine and clutter in an adaptive fashion.
Ground-penetrating radar signal processing for the detection of buried objects
Mitchell Walters,
Ephrahim Garcia
Show abstract
In this work the singular value decomposition (SVD) is used to analyze matrices of ground penetrating radar (GPR) data.
The targets to be detected are Russian PMN antipersonnel landmines and improvised explosive devices constructed from
155mm artillery shells. Target responses are simulated with GPRmax 2D, a simulation package based on the Finite-
Difference-Time-Domain method. First, the utility of the SVD for image enhancement and reconstruction is
demonstrated. Then the singular values and singular vectors of the decomposed matrices are analyzed with the goal of
finding properties that will aid in the development of automated underground detection algorithms.
Adaptive Gaussian mixture models for pre-screening in GPR data
Show abstract
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.
Physics-based features for identifying contextual factors affecting landmine detection with ground-penetrating radar
Show abstract
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.
Signal Processing Ground-Penetrating Radar Data II
Multiple instance learning framework for landmine detection using ground-penetrating radar
Show abstract
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.
Contextual learning in ground-penetrating radar data using Dirichlet process priors
Show abstract
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.
Exploiting spectral content for image segmentation in GPR data
Show abstract
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.
Comparative analysis of clutter suppression techniques for landmine detection using ground-penetrating radar
Show abstract
In this study, we provide an extensive comparison of different clutter suppression techniques that are proposed
to enhance ground penetrating radar (GPR) data. Unlike previous studies, we directly measure and present
the effect of these preprocessing algorithms on the detection performance. Basic linear prediction algorithm
is selected as the detection scheme and it is applied to real GPR data after applying each of the available
clutter suppression techniques. All methods are tested on an extensive data set of different surrogate mines
and other objects that are commonly encountered under the ground. Among several algorithms, singular value
decomposition based clutter suppression stands out with its superior performance and low computational cost,
which makes it practical to use in real-time applications.
Infrared
Detection of buried mines and explosive objects using dual-band thermal imagery
Show abstract
We demonstrate the development and use of novel image processing methods to combine dual-band (MWIR and LWIR)
images from SELEX GALILEO's Condor II camera to extract characteristics of observed scenes comprising buried
mines and explosive objects. We discuss the development of a statistical processing technique to extract the different
characteristics of the two bands. We further present a statistical classifier used to detect targets on independently trained
images with a high detection probability and low false negative rates and discuss methods to mitigate the impact of false
positives through the selective processing of image regions and the contextual interpretation of the scene content.
Investigation of the potential use of hyperspectral imaging for stand-off detection of person-borne IEDs
Show abstract
Advances in hyperspectral sensors and algorithms may benefit the detection of person-borne
improvised explosive devices (PB-IEDs). While portions of the electromagnetic spectrum, such
as the x-ray and terahertz regions, have been investigated for this application, the spectral region
of the ultraviolet (UV) through shortwave infrared (SWIR) (250 nm to 2500 nm) has received
little attention. The purpose of this work was to investigate what, if any, potential there may be
for exploiting the spectral region of the UV through SWIR for the detection of hidden objects
under the clothing of individuals. The optical properties of both common fabrics and materials
potentially used to contain threat objects were measured, and a simple example using a
hyperspectral imager is provided to illustrate the combined effect. The approach, measurement
methods, and results are described in this paper, and the potential for hyperspectral imaging is
addressed.
Characterizing optical properties of disturbed surface signatures
Show abstract
The burial of objects disturbs the ground surface in visually perceptible ways. This project investigated how such
information can inform detection via imaging from visible through mid-infrared wavelengths. Images of the ground
surface where objects were buried were collected at multiple visible through mid-infrared wavelengths prior to burial
and afterward at intervals spanning approximately two weeks. Signs of soil disturbed by emplacement change over time
and exposure in the natural environment and vary in salience across wavelengths for different time periods. Transient
cues related to soil moisture or illumination angle can make signatures extraordinarily salient under certain conditions.
Longpass shortwave infrared and multi-band mid-infrared imaging can enhance the signature of disturbed soils over
visible imaging. These findings add knowledge and understanding of how soil disturbances phenomena can be exploited
to aid detection.
Gaussian mixture models for measuring local change down-track in LWIR imagery for explosive hazard detection
Show abstract
Burying objects below the ground can potentially alter their thermal properties. Moreover, there is often soil disturbance
associated with recently buried objects. An intensity video frame image generated by an infrared camera in the medium
and long wavelengths often locally varies in the presence of buried explosive hazards. Our approach to automatically
detecting these anomalies is to estimate a background model of the image sequence. Pixel values that do not conform to
the background model may represent local changes in thermal or soil signature caused by buried objects. Herein, we
present a Gaussian mixture model-based technique to estimate the statistical model of background pixel values. The
background model is used to detect anomalous pixel values on the road while a vehicle is moving. Foreground pixel
confidence values are projected into the UTM coordinate system and a UTM confidence map is built. Different
operating levels are explored and the connected component algorithm is then used to extract islands that are subjected to
size, shape and orientation filters. We are currently using this approach as a feature in a larger multi-algorithm fusion
system. However, in this article we also present results for using this algorithm as a stand-alone detector algorithm in
order to further explore its value in detecting buried explosive hazards.
Signal Processing and Sensor Fusion
Detection of targets in forward-looking infrared imaging using a multiple instance learning framework
Show abstract
In this paper we describe a method for generating cues of targets present in the field-of-view of an infrared (IR) camera
installed on a moving vehicle. A typical two class classifier requires the building of an image library containing
manually extracted examples of both types of objects, i.e., targets and non-targets. This approach, usually tedious on
static images, becomes intractable when using video sequences taken from a moving vehicle. To avoid a detailed manual
segmentation of video sequences, we employ a multiple instance learning (MIL) framework that allows training a two
class classifier just by specifying if a frame contains targets or not. The proposed method has three steps. First, for each
frame of a training run, we generate a set of possible points of interest using a corner detection algorithm. Second, for
the same training run, we tag each frame as positive (target hits present) or negative (only non-target hits present). Each
hit is described using the local binary pattern (LPB) features computed around its image location. The generated LPB
feature vectors, together with their frame tag, are used by the MIL training framework to generate a set of target LPB
prototypes. Although many regular classifiers may be trained using MIL, in this paper we employed a simple approach
based on the nearest prototype. In the last step, we used the computed prototypes to classify the corner hits detected in
several test video sequences. To validate our approach, we present results obtained on several runs gathered with a long
wave infrared (LWIR) camera mounted on a moving vehicle.
Sensor fusion approaches for EMI and GPR-based subsurface threat identification
Show abstract
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.
Change detection for detecting potential threats in forward-looking video and downward-looking ground-penetrating radar data
Show abstract
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.
Algorithm fusion in forward-looking long-wave infrared imagery for buried explosive hazard detection
Show abstract
In this article, we propose a method to fuse multiple algorithms in a long wave infrared (LWIR) system in the context of
forward looking buried explosive hazard detection. A pre-screener is applied first, which is an ensemble of local RX
filters and mean shift clustering in UTM space. Hit correspondence is then performed with an algorithm based on corner
detection, local binary patterns (LBP), multiple instance learning (MIL) and mean shift clustering in UTM space. Next,
features from image chips are extracted from UTM confidence maps based on maximally stable extremal regions
(MSERs) and Gaussian mixture models (GMMs). These sources are then fused using an ordered weighted average
(OWA). While this fusion approach has yet to improve the overall positive detection rate in LWIR, we do demonstrate
false alarm reduction. Targets that are not detected by our system are also not detected by a human under visual
inspection. Experimental results are shown based on field data measurements from a US Army test site.
Validating spectral spatial detection based on MMPP formulation
Show abstract
Spectral, shape or texture features of the detected targets are used to model the likelihood of the targets to be
potential mines in a minefield. However, some potential mines can be false alarms due to the similarity of the mine
signatures with natural and other manmade clutter signatures. Therefore, in addition to the target features, spatial
distribution of the detected targets can be used to improve the minefield detection performance. In our recently
published SPIE paper, we evaluated minefield detection performance for both patterned and unpatterned minefields
in highly cluttered environments, simultaneously using both target features and target spatial distributions that define
Markov Marked Point Process (MMPP). The results have suggested that proper exploitation of spectral/shape
features and spatial distributions can indeed contribute improved performance of patterned and unpatterned
minefield detection. Also, the ability of the algorithm to detect the minefields in highly cluttered environments
shows the robustness of the developed minefield detection algorithm based on MMPP formulation. Moreover, the
results show that the MMPP minefield detection algorithm performs significantly better than the baseline algorithm
employing spatial point process with false alarm mitigation. Since these results were based on the simulated data, it
is not clear that the MMPP detection algorithm has fully achieved its best performance. To validate its
performance, an analytical solution for the minefield detection problem will be developed, and its performance will
be compared with the performance of the simulated solution. The analytical solution for the complete minefield
detection problem is intractable due to a large number of detections and the variation of the number of detected
mines in the minefield process. Therefore, an analytical solution for a simplified detection problem will be derived,
and its minefield performance will be compared with the minefield performance obtained from the simulation in the
same MMPP framework for different clutter rates.
Using predictive distributions to estimate uncertainty in classifying landmine targets
Show abstract
Typical classification models used for detection of buried landmines estimate a singular discriminative output. This
classification is based on a model or technique trained with a given set of training data available during system
development. Regardless of how well the technique performs when classifying objects that are 'similar' to the training
set, most models produce undesirable (and many times unpredictable) responses when presented with object classes
different from the training data. This can cause mines or other explosive objects to be misclassified as clutter, or false
alarms. Bayesian regression and classification models produce distributions as output, called the predictive distribution.
This paper will discuss predictive distributions and their application to characterizing uncertainty in the classification
decision, from the context of landmine detection. Specifically, experiments comparing the predictive variance produced
by relevance vector machines and Gaussian processes will be described. We demonstrate that predictive variance can be
used to determine the uncertainty of the model in classifying an object (i.e., the classifier will know when it's unable to
reliably classify an object). The experimental results suggest that degenerate covariance models (such as the relevance
vector machine) are not reliable in estimating the predictive variance. This necessitates the use of the Gaussian Process
in creating the predictive distribution.
Buried explosive hazard detection using forward-looking long-wave infrared imagery
Show abstract
Trainable size-contrast filters, similar to local dual-window RX anomaly detectors, utilizing the Bhattacharyya
distance are used to detect buried explosive hazards in forward-looking long-wave infrared imagery. The imagery,
captured from a moving vehicle, is geo-referenced, allowing projection of pixel coordinates into (UTM) Universal
Transverse Mercator coordinates. Size-contrast filter detections for a particular frame are projected into UTM
coordinates, and peaks are detected in the resulting density using the mean-shift algorithm. All peaks without a minimum
number of detections in their local neighborhood are discarded. Peaks from individual frames are then combined into a
single set of tentative hit locations, and the same mean-shift procedure is run on the resulting density. Peaks without a
minimum number of hit locations in their local neighborhood are removed. The remaining peaks are declared as target
locations. The mean-shift steps utilize both the spatial and temporal information in the imagery. Scoring is performed
using ground truth locations in UTM coordinates. The size-contrast filter and mean-shift parameters are learned using a
genetic algorithm which minimizes a multiobjective fitness function involving detection rate and false alarm rate.
Performance of the proposed algorithm is evaluated on multiple lanes from a recent collection at a US Army test site.