Forward looking ground penetrating radar (FLGPR) has the benefit of detecting objects at a significant standoff distance. The FLGPR signal is radiated over a large surface area and the radar signal return is often weak. Improving detection, especially for buried in road targets, while maintaining an acceptable false alarm rate remains to be a challenging task. Various kinds of features have been developed over the years to increase the FLGPR detection performance. This paper focuses on investigating the use of as many features as possible for detecting buried targets and uses the sequential feature selection technique to automatically choose the features that contribute most for improving performance. Experimental results using data collected at a government test site are presented.
In this paper, we develop an approach to detect explosive hazards designed to attack vehicles from the side of a
road, using a side looking synthetic aperture acoustic (SAA) sensor. This is done by first processing the raw data using a
back-projection algorithm to form images. Next, an RX prescreener creates a list of possible targets, each with a
designated confidence. Initial experiments are performed on libraries of the highest confidence hits for both target and
false alarm classes generated by the prescreener. Image chips are extracted using pixel locations derived from the target’s
easting and northing. Several feature types are calculated from each image chip, including: histogram of oriented
gradients (HOG), and generalized column projection features where the column aggregator takes the form of the
minimum, maximum, mean, median, mode, standard deviation, variance, and the one-dimensional fast Fourier transform
(FFT). A support vector machine (SVM) classifier is then utilized to evaluate feature type performance during training
and testing in order to determine whether the two classes are separable. This will be used to build an online detection
system for road-side explosive hazards.
This paper proposes a machine learning algorithm for subsurface object detection on multiple-input-multiple-output (MIMO) forward-looking ground-penetrating radar (FLGPR). By detecting hazards using FLGPR, standoff distances of up to tens of meters can be acquired, but this is at the degradation of performance due to high false alarm rates. The proposed system utilizes an anomaly detection prescreener to identify potential object locations. Alarm locations have multiple one-dimensional (ML) spectral features, two-dimensional (2D) spectral features, and log-Gabor statistic features extracted. The ability of these features to reduce the number of false alarms and increase the probability of detection is evaluated for both co-polarizations present in the Akela MIMO array. Classification is performed by a Support Vector Machine (SVM) with lane-based cross-validation for training and testing. Class imbalance and optimized SVM kernel parameters are considered during classifier training.
A sliding window based prescreening algorithm, utilizing multi-scale histogram of oriented gradient (MS-HOG) features and a linear support vector machine (SVM) classifier, for detection of buried explosive hazards in forward-looking infrared (FL-IR) and forward-looking ground penetrating radar (FL-GPR) data is presented. This algorithm is compared to previously published FL-IR and FL-GPR prescreening algorithms. The MS-HOG prescreening approach has higher computational complexity, but improves overall detection rates, especially for low-contrast and obscured target signatures. Results are presented on several data sets collected at US Army test sites. These collections span several days, and the FL-IR collections include imagery from both long-wave and mid-wave infrared cameras at multiple standoff distances captured at different hours of the day and different times of the year.
In this paper we present an approach that uses Genetic Programming (GP) to evolve novel feature extraction algorithms for greyscale images. Our motivation is to create an automated method of building new feature extraction algorithms for images that are competitive with commonly used human-engineered features, such as Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG). The evolved feature extraction algorithms are functions defined over the image space, and each produces a real-valued feature vector of variable length. Each evolved feature extractor breaks up the given image into a set of cells centered on every pixel, performs evolved operations on each cell, and then combines the results of those operations for every cell using an evolved operator. Using this method, the algorithm is flexible enough to reproduce both LBP and HOG features. The dataset we use to train and test our approach consists of a large number of pre-segmented image “chips” taken from a Forward Looking Infrared Imagery (FLIR) camera mounted on the hood of a moving vehicle. The goal is to classify each image chip as either containing or not containing a buried object. To this end, we define the fitness of a candidate solution as the cross-fold validation accuracy of the features generated by said candidate solution when used in conjunction with a Support Vector Machine (SVM) classifier. In order to validate our approach, we compare the classification accuracy of an SVM trained using our evolved features with the accuracy of an SVM trained using mainstream feature extraction algorithms, including LBP and HOG.
It is well-known that a pattern recognition system is only as good as the features it is built upon. In the fields of image processing and computer vision, we have numerous spatial domain and spatial-frequency domain features to extract characteristics of imagery according to its color, shape and texture. However, these approaches extract information across a local neighborhood, or region of interest, which for target detection contains both object(s) of interest and background (surrounding context). A goal of this research is to filter out as much task irrelevant information as possible, e.g., tire tracks, surface texture, etc., to allow a system to place more emphasis on image features in spatial regions that likely belong to the object(s) of interest. Herein, we outline a procedure coined soft feature extraction to refine the focus of spatial domain features. This idea is demonstrated in the context of an explosive hazards detection system using forward looking infrared imagery. We also investigate different ways to spatially contextualize and calculate mathematical features from shearlet filtered candidate image chips. Furthermore, we investigate localization strategies in relation to different ways of grouping image features to reduce the false alarm rate. Performance is explored in the context of receiver operating characteristic curves on data from a U.S. Army test site that contains multiple target and clutter types, burial depths, and times of day.
A convolutional neural network (CNN) approach to recognition of buried explosive hazards in forward-looking long-wave infrared (FL-LWIR) imagery is presented. The convolutional filters in the first layer of the network are learned in the frequency domain, making enforcement of zero-phase and zero-dc response characteristics much easier. The spatial domain representations of the filters are forced to have unit l2 norm, and penalty terms are added to the online gradient descent update to encourage orthonormality among the convolutional filters, as well smooth first and second order derivatives in the spatial domain. The impact of these modifications on the generalization performance of the CNN model is investigated. The CNN approach is compared to a second recognition algorithm utilizing shearlet and log-gabor decomposition of the image coupled with cell-structured feature extraction and support vector machine classification. Results are presented for multiple FL-LWIR data sets recently collected from US Army test sites. These data sets include vehicle position information allowing accurate transformation between image and world coordinates and realistic evaluation of detection and false alarm rates.
In this paper we describe an approach for optimizing the parameters of a Support Vector Machine (SVM) as part of an algorithm used to detect buried objects in forward looking infrared (FLIR) imagery captured by a camera installed on a moving vehicle. The overall algorithm consists of a spot-finding procedure (to look for potential targets) followed by the extraction of several features from the neighborhood of each spot. The features include local binary pattern (LBP) and histogram of oriented gradients (HOG) as these are good at detecting texture classes. Finally, we project and sum each hit into UTM space along with its confidence value (obtained from the SVM), producing a confidence map for ROC analysis. In this work, we use an Evolutionary Computation Algorithm (ECA) to optimize various parameters involved in the system, such as the combination of features used, parameters on the Canny edge detector, the SVM kernel, and various HOG and LBP parameters. To validate our approach, we compare results obtained from an SVM using parameters obtained through our ECA technique with those previously selected by hand through several iterations of “guess and check”.
There is a strong need to develop an automatic buried explosive hazards detection (EHD) system for purposes such as route clearance. In this article, we put forth a new automatic detection system, which consists of keypoint identification, feature extraction, classification and clustering. In particular, we focus on a new soft feature extraction process from forwardlooking long-wave infrared (FL-LWIR) imagery based on the use of an importance map derived from a bank of Gabor energy filters. Experiments are conducted using a variety of target types buried at varying depths at a U.S. Army test site. An uncooled LWIR camera is used and the collected data spans multiple lanes and times of day (due to diurnal temperature variation that occurs in IR). The preliminary receiver operating characteristic (ROC) curve-based performance presented herein is extremely encouraging for FL-EHD.
In previous work an automatic detection system for locating buried explosive hazards in forward-looking longwave infrared (FL-LWIR) and forward-looking ground penetrating radar (FL-GPR) data was presented. This system consists of an ensemble of trainable size-contrast filters prescreener coupled with a secondary classification step which extracts cell-structured image space features, such as local binary patterns (LBP), histogram of oriented gradients (HOG), and edge histogram descriptors (EHD), from multiple looks and classifies the resulting feature vectors using a support vector machine. Previously, this system performed image space to UTM coordinate mapping under a flat earth assumption. This limited its applicability to flat terrain and short standoff distances. This paper demonstrates a technique for dense 3D scene reconstruction from a single vehicle mounted FL-LWIR camera. This technique utilizes multiple views and standard stereo vision algorithms such as polar rectification and optimal correction. Results for the detection algorithm using this 3D scene reconstruction approach on data from recent collections at an arid US Army test site are presented. These results are compared to those obtained under the flat earth assumption, with special focus on rougher terrain and longer standoff distance than in previous experiments. The most recent collection also allowed comparison between uncooled and cooled FL-LWIR cameras for buried explosive hazard detection.
In this paper we investigate a new approach for representing objects in FLIR images based on shearlets. Similar to wavelets, shearlets represent an affine system for image representation obtained by scaling and translation of a generating function called mother shearlet. Unlike wavelets, the mother shearlet has an extra parameter called shear that allows the shearlet transform to be anisotropic. Anisotropic property of the shearlet transform could allow for a better representation of objects with irregular shape. We test our representation methodology on Froward looking long wave infrared (LWIR) images obtained from an IR camera installed on a moving vehicle. Objects of interest (spots) are detected in each frame using a prescreener presented in our previous work. Each spot is then represented using its shearlet features and assigned a confidence coming from a support vector machine classifier. We compare shearlets to various traditional features such as local binary patterns (LPB) and histogram of gradients (HOG). The comparison is performed on a large dataset that consists of 16 runs at a US Army test site.
This paper proposes an effective anomaly detection algorithm for forward-looking ground-penetrating radar (FLGPR).
The challenges in detecting explosive hazards with FLGPR are that there are multiple types of targets buried at different
depths in a highly-cluttered environment. A wide array of target and clutter signatures exist, which makes classifier
design difficult. Recent work in this application has focused on fusing the classifier results from multiple frequency subband
images. Each sub-band classifier is trained on suites of image features, such as histogram of oriented gradients
(HOG) and local binary patterns (LBP). This prior work fused the sub-band classifiers by, first, choosing the top-ranked
feature at each frequency sub-band in the training data and then accumulating the sub-band results in a confidence map.
We extend this idea by employing multiple kernel learning (MKL) for feature-level fusion. MKL fuses multiple sources
of information and/or kernels by learning the weights of a convex combination of kernel matrices. With this method, we
are able to utilize an entire suite of features for anomaly detection, not just the top-ranked feature. Using FLGPR data
collected at a US Army test site, we show that classifiers trained using MKL show better explosive hazard detection
capabilities than single-kernel methods.
In prior work, we describe multiple image space anomaly detection algorithms for the identification of buried
explosive materials in forward looking long wave infrared imagery. That work is extended here and focus is
placed on improved detection with respect to diurnal temperature variation. An ensemble of shape and size
independent image space anomaly detection algorithms are investigated. Specifically, anomalies are identified
according to change and blob detection. This anomaly evidence is aggregated and targets are found using an
ensemble of trainable size-contrast filters and weighted mean shift clustering. In addition, the blob detector
makes use of contrast-limited adaptive histogram equalization for image enhancement. Experimental results
are shown based on field data measurements from a U.S. Army test site.
Improvements to an automatic detection system for locating buried explosive hazards in forward-looking longwave
infrared (FL-LWIR) imagery, as well as the system's application to detection in confidence maps and forwardlooking
ground penetrating radar (FL-GPR) data, are discussed. The detection system, described in previous work,
utilizes an ensemble of trainable size-contrast filters and the mean-shift algorithm in Universal Transverse Mercator
(UTM) coordinates. Improvements of the raw detection algorithm include weighted mean-shift within the individual
size-contrast filters and a secondary classification step which exacts cell structured image space features, including local
binary patterns (LBP), histogram of oriented gradients (HOG), edge histogram descriptor (EHD), and maximally stable
extremal regions (MSER) segmentation based shape information, from one or more looks and classifies the resulting
feature vector using a support vector machine (SVM). FL-LWIR specific improvements include elimination of the need
for multiple models due to diurnal temperature variation. The improved algorithm is assessed on FL-LWIR and FL-GPR
data from recent collections at a US Army test site.
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.
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.
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.
This paper develops algorithms for the detection of interesting and abnormal objects in color and infrared imagery taken
from cameras mounted on a moving vehicle, observing a fixed scene. The primary purpose of detection is to cue a
human-in-the-loop detection system. Algorithms for direct detection and change detection are investigated, as well as
fusion of the two. Both methods use temporal information to reduce the number of false alarms.
The direct detection algorithm uses image self-similarity computed between local neighborhoods to determine interesting,
or unique, parts of an image. Neighborhood similarity is computed using Euclidean distance in CIELAB color space for
the color imagery, and Euclidean distance between grey levels in the infrared imagery. The change detection algorithm
uses the affine scale-invariant feature transform (ASIFT) to transform multiple background frames into the current image
space. Each transformed image is then compared to the current image, and the multiple outputs are fused to produce a
single difference image. Changes in lighting and contrast between the background run and the current run are adjusted
for in both color and infrared imagery. Frame-to-frame motion is modeled using a perspective transformation, the
parameters of which are computed using scale-invariant feature transform (SIFT) keypoint correspondences. This
information is used to perform temporal accumulation of single frame detections for both the direct detection and change
detection algorithms. Performance of the proposed algorithms is evaluated on multiple lanes from a data collection at a
US Army test site.
In this paper we describe a method for generating cues of possible abnormal objects present in the field of view of an
infrared (IR) camera installed on a moving vehicle. The proposed method has two steps. In the first step, for each frame,
we generate a set of possible points of interest using a corner detection algorithm. In the second step, the points related to
the background are discarded from the point set using an one class classifier (OCC) trained on features extracted from a
local neighborhood of each point. The advantage of using an OCC is that we do not need examples from the "abnormal
object" class to train the classifier. Instead, OCC is trained using corner points from images known to be abnormal object
free, i.e., that contain only background scenes. To further reduce the number of false alarms we use a temporal fusion
procedure: a region has to be detected as "interesting" in m out of n, m<n, consecutive frames in order to be reported as
abnormal. To choose the best classifier for our task, we compare the performance of three OCCs: nearest neighbor (OCNN),
SVM (OC-SVM) and Gaussian mixture (OC-GM). The comparison is performed using a set of about 900
background point neighborhoods for training and 400 for testing. The best performing OCC is then used to detect
abnormal objects in a set of IR video sequences obtained on a 1 mile long country road.
Rapid detection of landmines and explosive hazards is a critical issue for modern military operations. Due to the varied
nature of the objects of interest and the complexity of the surrounding, one approach is to utilize the superior recognition
capabilities of the human brain in the detection process. We are developing frameworks and algorithms to fuse image
data from multiple sources and to provide cuing capability for a human-in-the-loop detection system.
Forward-looking ground-penetrating radar (FLGPR) has received a significant amount of attention for use in explosive
hazards detection. A drawback to FLGPR is that it is sensitive to not only explosive hazards but also to benign objects,
which results in an excessive number of false detections. This paper presents our analysis of the explosive hazards
detection system developed by Planning Systems Inc (PSI). The PSI system combines FLGPR with an infrared (IR)
camera. We present an FLGPR target detection algorithm that leverages the multiple observations aspect of FLGPR.
The FLGPR target detections are then projected into the IR imagery. A Mahalanobis-metric classifier is then used to
reduce the number of false detections. We show that our proposed FLGPR target detection algorithm, coupled with our
IR-based false alarm reduction method, is effective at detecting explosive hazards while reducing the number of false
alarms.
This paper proposes a technique for using infrared (IR) imagery to eliminate false forward-looking ground penetrating
radar (FLGPR) detections by examining areas in IR images corresponding to FLGPR alarm locations. The FLGPR and
IR co-location is based on the assumption of a flat earth and the pinhole camera model. The parameters of the camera
and its location on the vehicle are not assumed to be known. The parameters of the model are estimated using a set of
correspondences gathered from the data utilizing the covariance matrix adaptation evolution strategy (CMA-ES)
optimization algorithm. Detection of false alarms is accomplished by generating a descriptor, consisting of various
statistics calculated from the IR images along with the FLGPR confidence value, for each alarm location. The alarms are
then classified based on the Mahalanobis distance between their descriptor and a multivariate normal distribution used to
model false alarms. The false alarm distribution is computed from training data where the validity of each alarm location
is already known. Using this technique, generally fifteen to twenty percent or more of the FLGPR false alarms can be
eliminated without losing any true alarms.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.