Automatic mine detection has recently become a subject of great importance to U.S. Navy. A number of approaches to this problem have been suggested so far. Current algorithms however do not provide sufficiently high performance results, especially in cases where mines are embedded in clutter. A thorough and fundamental understanding of target detection and recognition techniques is needed in order to significantly enhance the capabilities of automatic detection systems. We discuss here a possible approach to this problem, based on a theoretical model for image acquisition, that allows mine detection to be formulated as an inverse problem. New and near optimal algorithms may be developed as attempts to solving this problem.

The Coastal Systems Station (CSS) at Panama City, FL is developing an airborne multispectral sensor system which flies on an unmanned aerial vehicle for detecting mines in a coastal environment. This system is called the Coastal Battlefield Reconnaissance and Analysis (COBRA) system and has successfully completed preliminary developmental testing (DT-0). For this program, the Environmental Research Institute of Michigan (ERIM) developed a fieldable ground station including integrated aircraft tracking, real-time sensor data analysis, and a post processor testbed for developing and evaluating mine and minefield detection algorithms. A fully adaptive multispectral Constant False Alarm Rate mine detection algorithm was implemented in the post-processor by ERIM, along with patterned and scatterable minefield detection algorithms developed by CSS. The algorithms do not require prior knowledge of mine spectral signatures and thus are ideal for detecting a wide variety of mines with unknown or changing spectral signatures. COBRA DT-0 testing has been performed on actual minefields deployed at coastal and inland test sites. Preliminary results show that the COBRA system, coupled with these algorithms, meets the required minefield detection performance goals. This paper reviews the algorithm theory and implementation, overviews the ground station design, and presents minefield detection results from actual minefield imagery collected over realistic scenes during DT-0 testing.

The U.S. Marine Corps COBRA countermine surveillance program has developed, as a risk- reduction alternative, a near real-time processor for the output of the COBRA multispectral camera. This processor has been tested using approximately 13.5 hours of video data from the COBRA DT-0 developmental test, representing approximately 243,000 frames of multispectral data. The results have been very encouraging--the system is robust and the minefield detection performance has met the goals of the COBRA program. The MITRE COBRA prototype processor is built from commercial-off-the-shelf VME bus technology. Video capture is provided by a Transtech TDM 435 capture/display VME card. Control is performed on a GMSV64 Super Sparc card that resides in two VME slots. The compute engine consists of two Pentek 4270 Quad TMS320C40 digital signal processing boards. There are two additional 6U VME boards to provide fast SCSI IO. The system is capable of capturing, digitizing and processing the COBRA data stream at between one-eighth and one-half real-time, depending on processing options. The nominal compute power of the system is 2.2 GOPS, 450 MFLOPS. The system is easily upgradeable due to the open architecture--one proposed upgrade will be to increase the number of available TMS320C40 processors to sixteen, providing real-time performance without compromising the current investment in software and hardware. The software for the system is primarily written in C, with hand-optimized assembler code for portions of the compute kernel. The algorithm that is implemented is based on the MITRE minefield detection algorithm detailed at AeroSense '95. The system development required a registration algorithm--this was the only algorithm development that was performed, the rest of the algorithms coming from previous MITRE effort on the COBRA program. Lessons learned from the development and upgrade/test plans will be presented.

The detection of infrared (IR) targets immersed within an observed scene can be difficult when the target is embedded in a dominant clutter background. Multispectral IR imaging techniques for target detection have received increased attention over the past several years. This paper employs adaptive algorithms for reducing the effect of ground clutter in the presence of dependency, nonstationarity and system nonlinearities. Described are multispectral nonlinear adaptive algorithms as part of a detection scheme designed for small and/or dim point targets in IR images. Each spectral image captures a varying degree of target information. Individual scene observations and combined ones will be considered. Comparison between these frequency bands using nonlinear adaptive filters based on second order Volterra series expansion for multispectral imagery will be presented. Simulation results suggest that multispectral processing techniques have the potential for improving the detection of small and hard to find targets.

The surf zone environment represents a very difficult challenge for electro-optic surveillance programs. Data from these programs have been shown to contain dense clutter from vegetation, biological factors (fish), and man-made objects, and is further complicated by the water to land transition which has a significant impact on target signal-to-noise ratios (SNR). Also, targets can be geometrically warped from the sea surface and by occlusion from sand and breaking waves. The Program Executive Office Mine Warfare (PMO-210) recently sponsored a test under the Magic Lantern Adaptation (MLA) program to collect surf zone data. Analysis of the data revealed a dilemma for automatic target recognition algorithms; threshold target features high enough to reduce high false alarm rates from land clutter or low enough to detect and classify underwater targets. Land image typically have high SNR clutter with crisp edges while underwater images have lower SNR clutter with blurred edges. In an attempt to help distinguish between land and underwater images, target feature thresholds were made to vary as a function of the SNR of image features within images and as a function of a measure of the edge crispness of the image features. The feasibility of varying target feature thresholds to reduce false alarm rates was demonstrated on a target recognition program using a small set of MLA data. Four features were developed based on expected target shape and resolution: a contrast difference measure between circular targets and their local backgrounds, a signal-to-noise ratio, a normalized correlation, and a target circularity measure. Results showed a target probability of detection and classification (P_dc) of 50 - 78% with false alarms per frame of less than 4%.

We consider the problem of detecting minelike targets, imaged by means of multispectral sensors, that have been heavily corrupted by clutter. An effective detection approach needs to take into consideration the high correlation that is often present among bands in multispectral images and be robust against clutter. To this end, we here propose a two-step target detection approach. In particular, we first employ the Maximum Noise Fraction transform, in conjunction with vector-morphology, in order to reduce the effect of clutter and enhance the presence of targets. We then discuss a target detection algorithm, based on a morphological image reconstruction/marker fusion approach. We apply this algorithm to the problem of detecting minelike targets present in six-band aerial images, provided to us by the Coastal Systems Station, Naval Surface Warfare Center, Panama City, Florida. The proposed technique is relatively simple and requires only approximate knowledge of target size.

The detection of small targets in uncompressed imagery frequently incurs high computational cost due to area-based filtering and template matching processes. In particular, the convolution of a K-pixel filter with an N-pixel image typically requires work that is bounded below by O(KN). However, we have shown that such image-template operations can be computed in less than O(KN) time if the image is appropriately compressed. We call this technique compressive processing. In this two-art series of papers, we present supporting theory, derivations, and analyses of compressive image-template operations that frequently occur in automated target recognition practice. For example, compression ratios of 30:1 or greater have been reported for imagery when interframe differences are small. Similarly high compression ratios have been reported for video imagery using vector quantization (VQ) or visual pattern image coding (VPIC). We thus derive image operations such as edge detection and target classification that are applicable to VQ- and VPIC-compressed imagery, as well as to a VPIC-like transform, called Adaptive Vector Entropy Coding. In the case of edge detection and target classification over VQ- or VPIC-compressed imagery, we show that computational speedups of O(CR) can be obtained with appropriate data structural manipulation. For example, if VQ is employed with fixed-size, K-pixel encoding blocks, then edge detection can be achieved by entropy-based thresholding of the VQ codebook exemplars, at a cost of N/K block substitution operations. Given a codebook of size M vectors, an additional overhead of 2M comparisons may be required for validation purposes. A similar method is employed for VPIC, which encodes image patterns in terms of the encoding block mean, gradient intensity and orientation, and an index that references a bitmap pattern. In practice, the bitmap is derived from the encoding block's zero crossings about the block mean. Analyses emphasize performance measures such as computational cost, information loss, computational error, and compression ratio. Our algorithms are expressed in terms of image algebra, a rigorous, concise notation that unifies linear and nonlinear mathematics in the image domain. Since image algebra has been implemented on numerous sequential and parallel computers, our algorithms are feasible and widely portable.

Hierarchical neural network approaches have been developed first for combining high and low frequency (HF and LF) Side Scan Sonar imagery, and then for the combination of both acoustic images and Magnetic data. The adopted acoustic data fusion approach consists in a image-screening/HF, LF blob matching stage, followed by an information fusion/classification stage. Three variants of the information fusion/classification algorithm were conceived and evaluated based on `aggregate-feature-combining', `neural-network-discriminant-combining', and individual classifier `decision-based-combining', respectively. The `discriminant- combining' case yielded the best classification performance, and when compared with individual HF, LF classifier performance resulted in at least an order of magnitude reduction in the density of false alarms. Next, results are obtained for combining both acoustic and magnetic data using the described high and low frequency side scan sonar discriminant combining fusion algorithm as a starting point. In the next step, acoustic image pair `tokens' are associated with magnetic `tokens', resulting in three classes of resulting `tokens': `associated' acoustic-pair and magnetic tokens, isolated acoustic-pair tokens, and isolated magnetic `tokens'. Neural network output discriminants are derived for each of the three types of tokens mentioned above, and are employed to make classification decisions. The resulting Detection/Classification Algorithm is evaluated based on a combined ground truth obtained from both acoustic and magnetic sources.

An automatic, robust, adaptive clutter suppression, mine detection and classification processing string has been developed and applied to side-scan sonar imagery data. The overall processing string includes data pre-processing, adaptive clutter filtering (ACF), 2D normalization, detection, feature extraction, and classification processing blocks. The data pre-processing block contains automatic gain control and data decimation processing. The ACF technique designs a 2D adaptive range-crossrange linear FIR filter which is optimal in the Least Squares sense, simultaneously suppressing the background clutter while preserving an average peak target signature (normalized shape) computed a priori using training set data. A multiple reference ACF algorithm version was utilized to account for multiple target shapes (due to different mine types, multiple target aspect angles, etc.). The detection block consists of thresholding, clustering of exceedances and limiting their number, and a secondary thresholding process. Following feature extraction, the classification block applies a novel transformation to the data, which orthogonalizes the features and enables an efficient application of the optimal log-likelihood-ratio-test (LLRT) classification rule. The utility of the overall processing string was demonstrated with two side-scan sonar data sets. The ACF/feature orthogonalization based LLRT mine classification processing string provided average probability of correct mine classification and false alarm rate performance similar to that obtained when utilizing an expert sonar operator.

The shallow water coastal region is a particularly challenging environment in which to perform automated detection of mines. Sonar images taken in this region contain several types of acoustic image clutter including physical bottom clutter, high reverberation, multipath returns, and artifacts of sonar image construction such as high gain levels. The existence of image clutter is the primary difficulty in performing effective automated detection because this clutter typically results in a high false target rate. The identification and elimination of false targets is therefore required for effective automated target detection. A clutter recognition algorithm has been developed that has been very effective in reducing the false target rate for a set of side look sonars images. This algorithm acts as a supplement to detection and classification algorithms that identify and produce features spaces for possible minelike objects located in the images. The clutter recognition algorithm works by grouping detected objects by location and distance into image subspaces. Sets of group features are then calculated and the values of these features used to determine if each object is an actual mine or is part of an extended clutter formation. When used in conjunction with a statistics based detector and a fractal based target classifier, the clutter recognition algorithm produced an average false target rate of less than two false targets per image even though some of the more cluttered images contained over a hundred detected objects. Also, when used as an adjunct to a neural network classifier, the algorithm reduced the false target rate achieved by the network.

Detection of targets in sonar data is heavily reliant on a sparse number of clues: signal-to- noise ratio, size, and shape. Statistical analysis of the resolution cell energy distribution is an additional clue that examines a target size area for mean value, standard deviation, skewness, and kurtosis. Derivation and significance of mean and standard deviation are well known. The other two measures of frequency distribution, skewness, a measure of symmetrical distribution, and kurtosis, a measure of peakedness, are less well understood. The four measures were calculated for target size windows in digitized, beamformed sonar data. Statistical values for known targets were computed over successive pings. The combinations of these statistical measures provide an additional clue to target presence versus background. When used in conjunction with other known clues, these measures provide an increase in probability of detection/classification.

A modification to the hit/miss transform which has been developed to aid in the automatic detection of mines in sonar data is described. This modification allows the input thresholds for the data fed to the hit and the miss sections of the transform to be independently controlled. This makes it possible to have a high input threshold for the hit filter (this allows only high detections through) and a lower input threshold for the miss filter (this avoids splitting large but not as bright regions and thus reduces false alarms). The false alarms of particular interest are sand ridges and drag lines which must remain contiguous for the miss filter to properly exclude them. The modified hit/miss transform when fused with a matched filter has a similar detection rate and significantly lower false alarm rate when compared to the system which is presently in the field consisting of the matched filter fused with a neural network and a statistical analysis.

Detection processing of the Toroidal Volume Search Sonar beamformer output prior to image formation is used to increase the signal-to-reverberation. The energy detector and sliding matched filter perform adequately at close range but degrade considerably when the reverberation begins to dominate. The skewness matched filter offers some improvement. A dispersion based reconditioning algorithm, introduced in this paper, is shown to provide considerably improvement in the signal-to-reverberation at far range.

The feasibility of using residual (multiple stage) vector quantizer codevectors in a nearest neighbor classifier for direct classification of sonar pixel data is established. This approach combines the successive approximation process generated by the residual vector quantizer with successive decision making. Experimental results show that the probability of detection is about 80% and that the false alarm rate if about 5.6 false alarms per image. These initial performance benchmarks are encouraging considering the heuristic manner in which the residual vector quantizer codebooks were employed in the nearest neighbor classifier.

Time delay estimation has found applications in diverse fields such as underwater acoustic, radar, speech processing and biomedical signal processing. In underwater target detection in order to identify certain clues in the acoustic backscatter the time delays associated with the multi-specular returns must be estimated. In this paper a new joint time delay and signal estimation (JTDSE) procedure is proposed using the dyadic multi-resolution analysis framework. The multi-resolution analysis of a signal was performed through a discrete wavelet transform which is closely related to sub-band decomposition using filter banks. The goal of the proposed JTDSE scheme is to estimate the time delays corresponding to the multi-specular returns. Once the signals are decomposed, the time delays are estimated iteratively in each sub-band using two different adaptation mechanisms that minimize the sum of squared errors (MSE) between the reference and primary signals in the corresponding sub-band and scale. The localization of the minima of the MSE curves at different scales and sub-bands are used in order to arrive at the time delay estimates. Once the time delay estimates are validated among scales, the specular returns can then be separated from the backscattered signal for more accurate analysis of the residual part. Simulation results on synthesized as well as real data indicate the promise of this scheme for underwater target detection.

A new feature for the classification of echoes of a transmitted signal based on the generalized target description is described, and a framework for adaptive classification using it is presented. The generalized target description is a parametric model for the target impulse response. The feature is an order parameter from this model, which can be computed empirically from the growth rate of power as a function of a scale for a certain wavelet transform of the echo. A set of acoustic backscatter data consisting of returns from a mine and a rock with a linear FM transmit signal was analyzed. Parameters for the wavelet were computed from training sets so that this feature correctly distinguished 94% of the returns is test sets at 15 dB. The effectiveness of this feature as a classifier was found to degrade reasonably under increasing levels of synthetically generated reverberation noise. The simplest generalized target description model, a single order single center scatterer, was used. This model is not a realistic representation of the target impulse responses of either of the two objects, nevertheless it captured enough of the difference between the two to provide an effective classification tool.

A variety of experimental results indicate that Dolphins possess a unique and highly sophisticated sonar system. In addition, this sonar system is highly adaptive in detecting, discriminating and recognizing objects in highly reverberating and noisy environments. This paper presents possibly a new technique for target detection and recognition using the G- Transform and a new approach based on Resonance and Resonant Scattering Theory. These results show that this approach and signal processing technique used with neural networks may be useful in detection and identification of buried mine and minelike targets.

The problem of estimating multiple time delays in the presence of colored noise is considered in this paper. This problem is first converted to a frequency estimation problem by using the discrete Fourier transform. Then, sample lagged covariance matrices of the resulting signal are computed and studied in terms of their eigen-structure. These matrices are shown to be effective in extracting bases for the signal and noise subspace. MUSIC and Pencil-based frequency estimators are then derived using these subspaces. The effectiveness of the method is demonstrated on simulated backscatter which involves estimation of multiple specular components of the acoustic backscattered return from an underwater target.

No rapid (airborne) reconnaissance and survey capability exists to detect, localize, and discriminate buried and partially buried unexploded ordnance and ordnance and explosive waste. In the last decade, several factors point to the need for a new approach in site survey and cleanup methodology including an increasing number of areas being surplused by the Department of Defense (DOD), the continual public awareness of the location and status of formerly used defense sites, and the reduction in DOD budgets. The need for a safe, accurate, economical, and time-expedient land survey and detection capability has never been so great. Fused Airborne Sensor Technology (FAST) is derived from a similar proven multi-sensor fusion approach for mine reconnaissance against buried sea mines, the most difficult mines to detect in underwater environments. The FAST concept will use a single helicopter platform for spatial and temporal co-registration and fusion of sensor data. Sensors will include a superconducting magnetic field gradiometer, a two-color infrared camera, ground-penetrating radar, and a visible spectrum camera. The combination of these sensors allows for detection, identification, and localization of a wide spectrum of targets. Multiple `looks' at the target and associated clutter by sensors detecting different physical phenomena will allow lowering individual sensor thresholds to ensure each sensor detects all targets and clutter of interest. Sensor fusion will then be used to reject the clutter. Since sensor characteristic information is derivable through testing for both targets of interest and known clutter, a `fingerprinting' process can be used to pull the targets from the typically clutter-rich background. The FAST concept will be presented with sensor and fusion methodology for application to a wide variety of buried and partially buried targets.

The Improved Landmine Detector Concept Project was initiated in Autumn 1994 to develop a prototype vehicle mounted mine detector for low metal content and nonmetallic mines for a peacekeeping role on roads. The system will consist of a teleoperated vehicle carrying a highly sensitive electromagnetic induction (EMI) detector, an infrared imager (IR), ground probing radar (GPR), and a thermal neutron activation (TNA) detector for confirmation. The IR, EMI and TNA detectors have been under test since 1995 and the GPR will be received in June 1996. Results of performance trials of the individual detectors are discussed. Various design configurations and their tradeoffs are discussed. Fusion of data from the detectors to reduce false alarm rate and increase probability of detection, a key element to the success of the system, is discussed. An advanced development model of the system is expected to be complete by Spring 1997.

The US Army through the Night Vision and Electronic Sensors Directorate, Countermine Division is presently undertaking the application of multisensor methodologies to the mine detection need in a clutter-filled operational environment. Significant capabilities are being revealed through the conduct of US Government sponsored testing activities. This paper will discuss the state of multi-sensor applications to the US Army's Vehicular Mounted Mine Detection program. The intent is to present the community with results of developmental testing for multisensor mine detection approaches and define future strategies for technology integration of the Vehicular Mounted Mine Detection Advanced Technology Demonstration program.

This paper describes a vehicular mine detection system which automatically detects and marks buried land mines using very recent breakthroughs in mine detection technology. It combines proven detection technologies to realize a near 100% probability of detection for anti-tank mines. The detection system utilizes information from forward looking thermal and imaging system with downward looking radar and inductive pulse metal detection to sense the presence of buried or obscured land mines. Detection subsystems include: a sensor module, system processor, geographic location module and a real-time mine marking system. It also contains a registration process module which brings all selected targets, either in pixel space or sensor array position, to common platform coordinates, which in turn, are referenced to earth coordinates through GPS tracking of the platform. This registration process is extremely important, especially when integrating IR images from cameras whose positions are in non- nadir orientations. The detection system output provides geographic location of target mines, depth information, approximate mine shape and size, a natural scene image with graphically annotated mine locations.

A radar performance model is developed to analyze the capability of wide-band, high- resolution airborne synthetic aperture radar detection of shallow buried targets such as metallic M-20 mines and utility cables. Feasible target detection depths are estimated as functions of radar polarization, depression angle, and frequency in UHF and VHF bands. The performance model has incorporated wave propagation loss, due to wave attenuation inside the soil medium, wave reflection, and divergence at the air-ground interface, radar target cross section estimation via method of moments, radar interference, including both ambient man-made noise and empirical backscattered ground surface clutter, which is determined from existing clutter measurements for bare soil, rocks, and desert terrain.

Sensors incorporating superconducting quantum interference devices provide the greatest sensitivity for magnetic anomaly detection available with current technology. During the 1980s, the Coastal Systems Station (CSS) developed a superconducting magnetic gradiometer capable of operation outside of the laboratory environment. With this sensor, the CSS was able to demonstrate buried mine detection for the U.S. Navy. Subsequently, the sensor was incorporated into a multisensor suite onboard an underwater towed vehicle to provide a robust mine hunting capability for the Magnetic and Acoustic Detection of Mines Project. This sensor using thin film niobium and a new liquid helium cooling concept was developed to provide significant increases in sensitivity and detection range. In the late 1980s, a new class of `high- Tc' superconductor were discovered with critical temperatures above the boiling point of liquid nitrogen (77 K). This advance has opened up new opportunities for mine reconnaissance and hunting, especially for operation onboard small unmanned underwater vehicles. A high-Tc sensor concept using liquid nitrogen refrigeration has been developed and a test article of that concept is currently being evaluated for its applicability to mobile operation. The design principles for the two new sensor approaches and the results of their evaluations will be described. Finally, the implications of these advances to mine reconnaissance and hunting will be discussed.

Anti-personnel mines (APM) are spread over lots of countries which have been involved in wars. APMs usually do not possess self-destroying mechanisms and due to their long active time jeopardize the lives of millions of people. They are difficult to find with commercial metal detectors, because their content of metal is very low and in some cases even zero. Nevertheless more than 99% of all mines buried in the upper soil at least contain a little amount of metal inside the detonator. In this fact is founded the development of the surface scanning mine detection system ODIS (ordnance detection and identification system). The idea of ODIS is an imaging induction coil sensor (ICS) system which detects minimum metal volumes, calculates a real-time picture on a color screen showing the position, size and shape of metal parts buried in the ground and automatically marks the position on the ground. The measured data can be stored on disk and tape and can be post-processed for further information as depth, volume, kind of metal and contour in order to get a classification or even identification of the buried objects.

A gradiometer based on a microwave sapphire resonator is currently under development at the Applied Physics Laboratory. Its principle of operation is based on measurements of classical spring-proof mass displacements, which are converted to phase shifts in the output signal of a sapphire microwave resonator transducer. A prototype accelerometer used in the gradiometer has already demonstrated a figure-of-merit sensitivity [Q/(df/dx)] of 3.9 X 10² MHz/micrometers . The proposed design of the gradiometer we detail in this paper focuses on producing a simpler and more compact instrument. Results of a theoretical study on its ability to detect buried objects are also discussed. The problem of detection is challenging because of the instability of the downward continuation of a discrete set of surface gravity gradient measurements. Thus, detection capability depends on the size and density of the grid of measurements as well as the mass and depth of the buried object. Strategies for optimizing data collection and analysis are presented. Gradiometer is potentially useful in detecting buried objects such as mines since it is a totally passive measurement system which does not have to contact the ground directly. However, since the gravity gradient of a point mass decreases as a function of the cube of the distance, there is a strong advantage in having the instrument and the buried object as close to each other as possible.

Realistic field evaluations on the performance of hand-held mine detectors has until recently been limited to large bulky commercial test equipment instrumental with a network analyzer and computer controls. The Army's Night Vision and Electronic Sensors Directorate's Mine Detection Branch has just recently successfully completed field performance evaluations of new hand-held sensors at the Close-In Man Portable Mine Detector (CIMMD) Advanced Technology Demonstrations (ATD). Described in this paper will be one of the hand-held mine detector technologies evaluated at the CIMMD ATD and it's associated measured field performance results.

In a continuing effort to develop new sensor technologies for the detection of land mines and other UXO, a variety of plastic and metal mines were acquired for detection tests utilizing a passive millimeter wave sensor at 44 GHz and at 12 GHz. These inert mines were surface- laid, covered with dry leaves, or buried in sand or soil, and the resulting target scene was scanned from an overhead position using the single channel sensor, generating a 2D image of the minefield.

Real life sonar applications exist in which impulsive ocean channels tend to produce large- amplitude, short-duration interferences more frequently than Gaussian channels do. The stable law has been shown to successfully model noise over certain impulsive channels. In this paper, we propose new robust techniques for target detection and localization in the presence of noise modeled as a complex isotropic stable process. We develop optimal in the maximum likelihood sense approaches to the direction-of-arrival problem and we introduce the Cauchy Beamformer. We show that the Cauchy Beamformer provides better bearing estimates than the Gaussian Beamformer in a wide range of impulsive noise environments and for very low signal-to-noise ratio values. In addition, we derive the Cramer-Ratio bound on the estimation error covariance for the case of deterministic incoming signals retrieved in the presence of additive complex Cauchy noise. Finally, we demonstrate the robustness of the Cauchy Beamformer via simulation experiments.

A limited set of mine detection trials, using a 94 GHz Doppler Beamsharpening data gathering radar, are described. Measurements were made of arrays of HB876 anti-tank mines located on a tarmac runway and on grass. High spatial resolution radar images are presented with mines present and absent. The clutter statistics are summarized, and a method of thresholding the data using an Ordered Statistic Constant False Alarm Rate detection algorithm is described. The mine detection results are compared with theoretical predictions and it is shown that mines were clearly detected against the runway at the measurement range of 750 m. The poor detection results against the grass are explained in terms of radar speckle and signal to clutter ratio. Radar design techniques to overcome these limitations are described.

Random noise polarimetry is a new radar technique for high-resolution probing of subsurface objects and interfaces. Detection of buried targets is accomplished by cross-correlating the reflected signal by a time-delayed replica of the transmitted waveform. A unique signal processing scheme is used to inject coherent in the system to permit extraction of the wideband polarimetric scattering response of the buried object. This facilitates computation of the Stokes matrices of the target response which enhances the detection and identification process. Random noise polarimetry also possesses additional desirable features for subsurface probing such as immunity from detection and jamming. The paper discusses the theoretical foundations of random noise polarimetry and presents data acquired from various targets using a 1 - 2 GHz radar system fabricated by the University of Nebraska under contract to the U.S. Army Waterways Experiment Station. In addition, various signal processing algorithms used to analyze the polarimetric data are presented.

Images of buried ordnance can accelerate remediation through identification. This paper presents images of a buried, inert projectile. The images are plan views, at fixed but variable depths. The images were formed by processing measured reflectance through Fourier transformation, backward propagation, and inverse transformation. Data were measured in two tests. Both tests utilized a towed array of seven antennas. One test, in 1995, used frequencies between 187.5 and 487.5 MHz; the best images were from the 387.5 MHz data. An earlier test, in 1994, used frequencies 200, 350, and 500 MHz; the best images were formed from the 500 MHz data. The procedures for the two sets of data differed in relative orientation of the sensor antennas and projectile; in addition, soil dielectric constant values differed. Image displays also differed in image data interpolation.

Mine Countermeasures (MCM), an element of Mine Warfare, is a critical component of the US National Defense picture. MCM operations are mandatory since it is a simple and relatively inexpensive matter for opponents to place vast quantities of land and sea mines. Representation of MCM activities in simulations can substantially accelerate and optimize material and tactical means development to address this problem. A systematic effort to model and simulate the breadth of Mine Warfare operations is described. Developing computationally tractable models for inclusion in either time stepped or event driven MCM simulations is a formidable problem. Discussion of a particular strategy for dealing with modeling fidelity issues is provided, as are the issues and challenges associated with detection, recognition and system effectiveness representation.

The detection and disposal of anti-personnel landmines is one of the most difficult and intractable problems faced in ground conflict. This paper first presents current detection methods which use a separated aperture microwave sensor and an artificial neural-network pattern classifier. Several data-specific pre-processing methods are developed to enhance neural-network learning. In addition, a generalized Karhunen-Loeve transform and the eigenspace separation transform are used to perform data reduction and reduce network complexity. Highly favorable results have been obtained using the above methods in conjunction with a feedforward neural network. Secondly, a very promising idea relating to future research is proposed that uses acoustic modulation of the microwave signal to provide an additional independent feature to the input of the neural network. The expectation is that near-perfect mine detection will be possible with this proposed system.

The detection of spatial structure in fields of small targets is a challenging problem in high- level vision that hitherto has not been solved effectively. Although regular arrays of targets can be detected using local density and nearest-neighbor position estimators, such methods generally do not describe target field structure at multiple resolution levels. In this paper, we present several multiresolution techniques for detecting regular spatial structure in target field imagery. Our methods are tolerant of moderate deviations in target position, i.e., are amenable to irregular target placement, which is an important advantage when analyzing positions of targets deployed from slowly moving vehicles. We initially discuss local density computations based on pyramid-structured image representations, then progress to chord-based determination of target field structure. Methods for computing chord distribution parameters from histograms are presented that can be used in model-based detection of spatial structure in various target array configurations. Additionally, we discuss work-in-progress pertaining to data fusion methodologies for determining target field placement and structure in a variety of field scenarios. Analyses emphasize computational cost and error, as well as space requirements for parallel processing applications.

In this paper we explore the use of high-resolution array processing techniques for the detection and localization of buried objects with known shapes. A ground penetrating radar measurement geometry is considered where the scattered electric field due to an incident plane wave is observed over a linear receiver array positioned above the targets. We have modified the high-resolution algorithm MUSIC to explicitly account for the near-field physics and solved the target localization problem in two ways. The first method uses the MUSIC algorithm in a matched field processing scheme and determines both the bearing and the range of the targets. Under the second approach, the sensor array is divided into non-overlapping subarrays. The targets are assumed to lie in the far-field of each subarray ensuring a plane wave incidence over each partition. The DOAs are then found with the conventional plane wave MUSIC and the locations of the targets are determined by triangulation of the DOAs. Using simulated data we demonstrate that these techniques are quite useful for the detection and localization of metallic and dielectric mines as well as buried metallic drums. The favorable detection results are shown to hold over a wide range of soil conditions and signal to noise ratios.

The Navy and the Marine Corps have been continually concerned about the antivehicle and antiship mines. The development of effective minefield detection procedures are of great importance as they will enhance the ability of the Navy and the Marine Corps to perform their tasks safely. The need for effective minefield detection procedures based on tests of randomness have been addressed by navy researchers. In this article we present a survey of recent results in the area of one and two dimensional discrete scan statistics and discuss the relevance of these results to minefield detection. The testing procedures based on the discrete scan statistics rely heavily on accurate approximations of their distributions. A brief survey of the methods employed to arrive at these approximations is presented. Numerical results illustrating the use of the testing procedures based on the discrete scan statistics are given. Directions for future research are presented as well.

The purpose of this simulation is to compare the backscattered coefficients (or backscattered signals) among four different explosives (TNT, COMP B, TETRATOL, and PICRATOL) and two materials (STEEL and ALUMINUM). All six will be independently buried under four different sources: dry soil, and then moist soil independently covered with snow, with ice and finally with water. The four explosives were selected to represent non-metallic mines and STEEL and ALUMINUM were selected to represent metallic mines. The backscattered coefficient of a mine is calculated using the concept of pulse reflections and transmissions at junctions and pulse attenuations in the materials. A short pulse with a constant, initial voltage progresses through four junctions of different materials such as Antenna and Air, Air and Snow, Snow and Soil, and Soil and Mine with resulting pulses reflected, transmitted at these junctions and pulse attenuated in these materials. The product of the transmission ((tau) ) and reflection ((rho) ) coefficients ((tau) ₁(tau) ₂(tau) ₃(rho) ₄(tau) _3'(tau) _2'(tau) _1'), the attenuation factors (e^-2(alpha _airl_air)e^-2(alpha _soill_soil)), and the initial voltage is the backscattered signal from the mine. The backscattered coefficient of the mine is the ratio of the final voltage to the initial voltage in decibel.

Preliminary analysis has shown that accurate mine simulants are not available for metallic and non-metallic mine detectors; there is a need for a precise mine simulant. This paper contains the methodology for the development of a fuze simulant to replace the fuze in a real mine. Conclusions shall discuss the feasibility of an accurate fuze simulant. Mine detectors have been proven worthy by testing its capability to detect simulants or real mines without the fuze. Different detectors perceive changes in different properties of the mine; unfortunately, one simulant which reflects every property of the mine is not available. Another approach is to bury real mines, but for safety concerns, bury the mine without the fuze. This target still does not represent a real mine, especially for a metal detector. Many non-metallic mines contain a considerably quantity of metal in the fuze in comparison to the rest of the body of the mine; therefore, the absence of the fuze results in the absence of a significant amount of metal. The NVESD Countermine Division with the Energetic Materials Research Test Center of Socorro, NM, is developing a fuze simulant which more accurately represents the quantity of metal in a real time. They are investigating the contents of the fuze; this will be followed by fabrication and testing using infrared, ground penetrating radar, and metal detectors. The paper contains the methodology for the development of the fuze and discussions with respect to the benefits to mine detection.

We analyzed a time series of high resolution 8 - 13.4 micrometers scanner images of a sandbox with buried (3 - 10 cm depth) and unburied, metal and plastic AP and AT mines, surrogates and other targets. With a high resolution DUDA scanner operating in 8 - 13.4 micrometers all the surface laid targets were visible during the whole diurnal cycle. The buried targets were only visible during sunrise and sunset. The emissivity of the targets and sand could not be derived from the measured apparent temperatures and contact temperatures.

During the past decade, developments in lasers, gated intensified cameras, and high speed data processors have made it possible for electro-optic systems to play a key role in mine countermeasures. These developments have made it practical to search for sea and land mines in area where other physical observations were extremely difficult. Not only do these systems provide search capabilities, the data collected can provide significant insights into the optic properties of the environment. With increasing capability of optical equipment, new challenges lay ahead, providing for new ideas and research opportunities.

Experiments were conducted to determine if buried mines could be detected by measuring the change in reflectance spectra of vegetation above mine burial sites. Mines were laid using hand methods and simulated mechanical methods and spectral images were obtained over a three month period using a casi hyperspectral imager scanned from a personnel lift. Mines were not detectable by measurement of the shift of the red edge of vegetative spectra. By calculating the linear correlation coefficient image, some mines in light vegetative cover (grass, grass/blueberries) were apparently detected, but mines buried in heavy vegetation cover (deep ferns) were not detectable. Due to problems with ground truthing, accurate probabilities of detection and false alarm rates were not obtained.

In order to reduce the serious problem associated with the mining of important supply/communication roads by hostile parties during peacekeeping operations, the Canadian Department of National Defense has recently begun the development of a multi-sensor teleoperated mine detection vehicle, the Improved Landmine Detection Capability. One sensor identified as a serious candidate for that project is a passive IR camera. In the past, many organizations have assessed the efficiency of this technique of detection and reported widely fluctuating results. It is believed that the main reason for these fluctuations is associated with the ad hoc interpretations used by different researchers. In this paper, a more systematic analysis is presented which takes into account variables such as time of the day, time of the year, weather conditions, type of road and many others. A working model is proposed in order to facilitate the prediction of the IR signature of the buried land-mine and is compared with data acquired from multiple trials. These trials were done with live mines (without fuzes) and surrogates buried in different types of road (packed gravel and sand) and during different times of the day and different times of the year.

The measurement and removal of noise from images created using lateral migration backscatter radiography (LMBR) a form of Compton backscatter imaging (CBI) is applied to the detection and identification of landmines. The photons that interact with the landmine produce the signal component of interest. The signal is corrupted by both quantum and structured noise. The structured noise is due to photon interaction with non-mine material. Due to the strong response of all detectors to soil surface features and other buried objects, image enhancement methods are essential for landmine identification. A four detector system is used to generate the LMBR/CB images. The inner two detectors are uncollimated and positioned to optimally detect first scattered photons. The outer detectors are collimated to detect photons that have had two or more scatterings. The difference between the collimated and uncollimated detector responses to the different types of landmine image masking phenomena, form the basis of the image enhancement and landmine identification procedures. The surface feature information is obtained by the uncollimated detectors. The collimated detector signal contains information about the surface features as well as the buried objects. Using images from these two sets of detectors the surface objects can be analyzed for possible landmines and then removed. The buried objects can then be resolved. The measurements and image enhancements demonstrate that it is possible to detect 12' plastic landmines at a buried of 3' under simulated battlefield conditions.

The Compton Backscatter Imaging (CBI) technique has been applied successfully to detect buried plastic anti-tank landmines. The images acquired by a CBI system are often cluttered by surface features. Additionally, some buried objects give the same response as the plastic landmines. The landmine detection can be successful only when the detection system is capable of distinguishing between surface features and the mine-like objects. This can be accomplished by designing detectors that differentiate between the surface features and the buried objects. An understanding of the physical phenomena underlining the CB image formation helps us to design these detectors. To study the physics of the Compton backscattering, the photon transport in a CBI system is simulated using Monte-Carlo calculations with the generalized particle transport program MCNP. The photon tracks are graphically displayed using a visualization program SABRINA. On the basis of the results from these Monte-Carlo analyses, a four-detector system has been designed. This detector design utilizes the unique nature of various collision components of the scattered photons to generate separate images of buried objects and surface features. The success of this detector design is demonstrated through a series of analytically generated images. The results of the experimental measurements that validate these analytical predictions are brought out in a separate paper to be presented in this conference.

The questions of detector synthesis on the basis of the generalized signal processing algorithm in detection systems for mines and minelike targets are considered. The sensitivity to changing in the internal noise variance relative to the calculated variance and the probability of error for analog-to-digital conversion of input random process are investigated for the generalized and optimal detectors in digital detection systems for mines and minelike targets.

The image-based airborne detection of submerged threats and obstructions is frequently confounded by interfacial refraction and in-water scattering, which reduce image contrast and resolution. In particular, refractive effects spatially dissociate the received image, often to the extent that target shape is obscured. In the late 1980s, Schippnick addressed the problem of predicting deformations in trans-interfacial imagery. Over the past six years, we have developed models that simulate image degradation by trans-interfacial viewing. Additionally, our models can approximately reconstruct a submerged target given partial knowledge of sea surface topography and ocean optical properties. In this paper, we summarize recent advances in our simulation techniques that render our models more realistic physically, as well as amenable to processing SIMD-parallel architectures. We emphasize the simulation of image deformations, which facilitates algorithm development in support of image clarification or automated target recognition. In particular, we discuss requirements for the simulation of receiver noise, electronic effects (e.g., automatic gain control, saturation, and thermal noise), as well as spatial noise such as pixel blooming and point-spread effects resulting from image intensifier tubes. Additional emphasis is given to techniques for simulating caustic effects within the water column, including backscattering of intense irradiance from turbid water. The ability of our models to accommodate layered media is also considered. Analyses pertain to practical simulation issues such as computational cost, numerical error in predicting the spatial location of target features, and approximate computation of simulation operations. Our models are structured in terms of functional modules that correspond to individual physical processes and are coded in standard computer languages. Thus, our models are easily maintained and are portable to a wide variety of computers.

In this paper we consider the detection of small targets within an IR scene that are embedded in a dominant clutter background. A feasible approach toward addressing this issue is to invoke some form of signal processing that allows the clutter to be reduced from the scene prior to target detection. An NRLS scheme is employed which functions as whitening filter prior to matched filtering. The NRLS scheme is based on a second order truncated Volterra series expansion. The goal is to adapt to image nonstationarities and to equalize unknown system nonlinearities, prior to matched filtering. Simulation results based on both synthesized and `real world' nonstationary IR images are presented.

Acoustic detection of mine like objects in a shallow water environment is often limited by reverberation. This paper describes a toroidal array beamforming methodology for a shallow water environment, whereby vehicle attitude data is used to adaptively stabilize and scan several narrow sonar beams down a water column. By averaging over several beams the reverberation level is reduced, thus increasing the signal to noise ratio and greatly improving the area coverage rate for the sonar.

GEO-CENTERS has developed a ground mobile mine detection capability. It consists of a robust and viable multisensor vehicular testbed which can be deployed for integration and testing of current and future sensor technology as appropriate. The Integrated Ground Mobile Mine Detection Testbed (IGMMDT) system combines the capabilities of several sensors of diverse technologies, each of which has shown promise in mine detection, to provide an integrated approach to the problem in which the sensors supplement each other capabilities. The result is a broadening of the environmental and spatial ranges of detection, plus an increase in the speed of detection, location and area coverage. Integration of multiple sensors in an application of this type presents several predictable problems, not the least of which is the scoring, weighting and correlation of data from different sensors. Defining a common parameter set within which to integrate the sensor outputs and interpret the results is the primary problem. To get to this level, each sensor's output must be transformed into this common parameter set, which essentially means supplementing each sensor's output with enough additional processing as to make it an independent `smart sensor'. IGMMDT provides a convenient platform on which to develop integration techniques and processing methodologies usable in continuing sensor fusion.

We analyzed a time series of 94 GHz radiometer images of a sandbox with buried and unburied, metal and plastic AP and AT dummy mines. The images covered almost a complete 24 hour cycle, with both clear sky and rain conditions occurring. The AP nor the buried mines were visible at any time. The contrast of the visible mines depended mostly on the sky conditions. The plastic mines were (almost) always brighter than the background, the metal mines darker.

The Coastal Systems Station of the Naval Surface Warfare Center-Dahlgren Division is developing the Coastal Battlefield Reconnaissance and Analysis (COBRA) system. COBRA is a U.S. Marine Corps advanced technology demonstration utilizing multispectral video sensors deployed in an unmanned aerial vehicle (UAV) for automated minefield detection and location from the surfzone to inland areas. The system automatically encodes the craft attitude and position into video down-link for real-time ground tracking on a map, satellite image, or aerial photograph. The system has been developed and deployed extensively in a Cessna 172 during developmental testing (DT-0) and will undergo preliminary operation testing (OT-0) in a Pioneer UAV in the summer of 1996. This paper reviews the status and results of the COBRA developmental and operational testing.

This paper illustrates various issues involved in the use of imagery in mine warfare. Particular emphasis is given to environmental factors which can create false alarms.