The recognition of targets in infrared scenes is complicated by the wide variety of appearances associated with different thermodynamic states. We represent the variability in the thermodynamic signatures of targets via an expansion in terms of 'eigentanks' derived from a principal component analysis performed over the target's surface. Employing a Poisson sensor likelihood, or equivalently a likelihood based on Csiszar's I-divergence, a natural discrepancy measure for nonnegative images, yields a coupled set of nonlinear equations which must be solved to computed maximum a posteriori estimates of the thermodynamic expansion coefficients. We propose a weighted least-squares approximation to the Poisson loglikelihood for which the MAP estimates are solutions of linear equations. Bayesian model order estimation techniques are employed to choose the number of coefficients; this prevents target models with numerous eigentanks in their representation from having an unfair advantage over simple target models. The Bayesian integral is approximated by Schwarz's application of Laplace's method of integration; this technique is closely related to Rissanen's minimum description length and Wallace's minimum message length criteria. Our implementation of these techniques on Silicon Graphics computers exploits the flexible nature of their rendering engines. The implementation is illustrated in estimating the orientation of a tank and the optimum number of representative eigentanks for real data provided by the U.S. Army Night Vision and Electronic Sensors Directorate.

In our earlier work, we focused on pose estimation of ground- based targets as viewed via forward-looking passive infrared (FLIR) systems and laser radar (LADAR) imaging sensors. In this paper, we will study individual and joint sensor performance to provide a more complete understanding of our sensor suite. We will also study the addition of a high range- resolution radar (HRR). Data from these three sensors are simulated using CAD models for the targets of interest in conjunction with XPATCH range radar simulation software, Silicon Graphics workstations and the PRISM infrared simulation package. Using a Lie Group representation of the orientation space and a Bayesian estimation framework, we quantitatively examine both pose-dependent variations in performance, and the relative performance of the aforementioned sensors via mean squared error analysis. Using the Hilbert-Schmidt norm as an error metric, the minimum mean squared error (MMSE) estimator is reviewed and mean squared error (MSE) performance analysis is presented. Results of simulations are presented and discussed. In our simulations, FLIR and HRR sensitivities were characterized by their respective signal-to-noise ratios (SNRs) and the LADAR by its carrier-to-noise ratio (CNR). These figures-of-merit can, in turn, be related to the sensor, atmosphere, and target parameters for scenarios of interest.

We analyze and assess the underlying assumptions and characteristics of the standard Gaussian automatic target recognition algorithm. An analysis of the theoretical formulation of the basic algorithm is carried out in which the important assumption of Gaussian multivariate feature distribution is replaced with the assumption of generalized Rayleigh multivariate feature distribution. Closed form analytical expressions are worked out for the associated characteristic and detection probability functions. Numerical analysis of the results is performed which reveals that superior performance characteristics can arise in the generalized Rayleigh distribution-based case. An empirical analysis of a computer programmatic implementation of the basic Gaussian algorithm is also carried out to explore the sensitivity of the generated numerical results to the variation of those parameters which are intrinsic to the code. It is explicitly demonstrated that the statistics of the receiver-operator characteristics yielded by the code are extremely sensitive to this set of parameters, and that this sensitivity can lead to potentially ambiguous results in important cases.

This paper describes the impact of proper site characterization and ground truthing on test results for the detection of Unexploded Ordnance (UXO). Techniques employed to do proper site characterization and methods used to generate accurate and meaningful ground truth data are presented within the context of recent data collections held at the Joint UXO Coordination Office's (JUXOCO) test site at Ft. AP Hill VA. The risks associated with poor site preparation or inaccurate ground truthing are presented with illustrative examples based upon past data collection exercises. Key documentation including collection plans and data formats, as well as Web- based data distribution and coordination approaches developed and in use by the Night Vision and Electronic Sensors Directorate (NVESD) in support of the JUXOCO mission are highlighted. Similarities between the UXO mission and the Countermine mission are used to extend the approaches recommended in this paper to related Countermine data collection exercises. The techniques discussed within this paper constitute a key component of a comprehensive Aided Target Recognition (ATR) test and evaluation methodology used by NVESD and made available to the ATR community to further algorithm development and maturity within the industry.

One of the problems with taking aided target recognizer software written for a uniprocessor and hosting it on a multiprocessor running in real time is determining the allocation of data and functions across multiple processors. The partitioning varies based on the algorithm and the computer architecture. Often, more than one partitioning is possible. Ideally, one would like to evaluate the possible partitioning, choose a one that meets requirements, and modify or rewrite the existing program to embody that partitioning. The paper describes the steps that were used to rehost a real aided target recognition algorithm onto a commercial off-the- shelf embedded multiprocessor, with an emphasis on the use of performance modeling to determine which partitioning schemes will result in code which meets requirements. Libraries that facilitate the transition from uniprocessor to multiprocessor are also discussed; the libraries are being evaluated as part of an NVESD effort for the High Performance Computing Modernization Office.

The preprocessing or detection stage of an automatic target recognition system extracts areas containing potential targets from a battlefield scene. These potential target images are then sent to the classification stage to determine the identity of the targets. It is highly desirable at the preprocessing stage to minimize incorrect rejection rate. This, however, results in a high false alarm rate. In this paper, we present a new technique to reject false alarms (clutter images) produced by the preprocessing stage. Our technique, region-based principal component analysis (PCA), uses topological features of the targets to reject false alarms. A potential target is divided into several regions and a PCA is performed on each region to extract regional feature vectors. We propose to use regional feature vectors of arbitrary shapes and dimensions that are optimized for the topology of a target in a particular region. These regional feature vectors are then used by a two-class classifier based on the learning vector quantization to decide whether a potential target is a false alarm or a real target.

An effective clutter rejection scheme is needed to distinguish between clutter and targets in a high-performance automatic target recognition system. In this paper, we present a clutter rejection scheme that consists of an eigenspace transformation and a multilayer perceptron (MLP). The input to the clutter rejector module is the output of the detector that provides the potential regions (target chips). We first use an eigen transformation for feature extraction and dimensionality reduction. The transformations considered in this research are principal component analysis (PCA) and the eigenspace separation transform (EST). These transformations differ in their capabilities to enhance the class separability and to compact the information (energy) for a given training set. The result of the eigenspace transformation is then fed to an MLP that predicts the identity of the input, which is either a target or clutter. To search for the optimal performance, we use different sets of eigentargets and construct the matching MLPs. Modified from the popular Qprop algorithm, we devise an MLP training algorithm that seeks to maximize the class separation at a given false-alarm rate, which does not necessarily minimize the average deviation of the MLP outputs from their target values. Experimental results are presented on a huge and realistic data set of forward-looking infrared (FLIR) imagery.

Abstract not available.

This paper deals with a Ph-D work about Automatic Target Recognition in Infrared aerial image sequences. The targets to be recognized are ground military vehicles like tanks or lorries. . . During the first step of the Automatic Target Recognition system simulation, the targets are segmented and tracked using an innovative active contour model. The active contour is based on snakes, robust statistics and it uses temporal information on the deformation of the target, such information being acquired during the sequence. This is performed in order to improve the tracking and the recognition to follow. The second step of the ATR system is the on-line recognition of the tracked and segmented objects. To that end, we use two modules based on pre-trained artificial neural networks. One is dedicated to target classification, the other to target identification. Both receive as input the Fourier descriptor of the extracted target shape. This method is validated both on Air-To-Ground IR seeker images and Ground IR camera images.

In this paper, the structural risk minimization (SRM) criterion is employed to train a large margin classifier, the support vector machine (SVM). Its relative performance is compared with traditional classifiers employing hyperplanes against a realistic difficult problem, the synthetic aperture radar (SAR) automatic target recognition (ATR). In most pattern recognition applications, the task is to perform classification into a fixed number of classes. However, in some practical cases, such as ATR, one also needs to carry out a reliable pattern rejection. Experimental results showed that the SVM with the Gaussian kernels performs well in target recognition. Moreover, the SVM is able to form a local or 'bounded' decision region that presents better rejection to confusers.

This paper proposes the minimization of a new cost function while training backpropagation (BP) neural networks to solve pattern classification problems. The new cost function is referred to as the gain-weighted normalized-target mean-square error (GWNTMSE). The paper proves that the minimization of the GWNTMSE is optimal in the sense of yielding network classifier with minimum variance from the optimal Bayes classifier in the limit of an asymptotically large number of statistically independent training patterns. Experimental results are presented. The application selected is the classification of ship targets in airborne synthetic aperture radar (SAR) imagery. The number of ship classes is 8. They represent 2 destroyers, 2 cruisers, 2 aircraft carries, a frigate, and a support ship. The obtained results indicate that BP classifiers trained by minimizing the GWNTMSE consistently outperform those trained by minimizing the standard MSE.

This paper presents a new approach for the classification of SAR targets that combines maximally decimated directional filter banks with higher-order neural networks (HONNs). HONNs are neural networks that permit the input signals to be multiplied together in addition to the more common operations such as weighting, summing, and pointwise nonlinearities of typical neural nets. HONNs have long been proposed as image classifiers whose performance can be made invariant to geometric transformations of the input imagery by using a method for decreasing dimensionality such as coarse coding. Most past image classifiers using HONNs have been tuned for carefully thresholded binary images, which generally cannot be derived from low-contrast imagery such as SAR without a significant loss of information. As an alternative, we use a novel HONN implementation that accepts gray-level input pixels using directional filter banks. In order to do this, a new modified tree-structured directional filter bank structure is proposed in this paper, where each of the subbands has directional visual information from a given input. The performance of the proposed approach is demonstrated with imagery taken from the public MSTAR database.

Metal and nonmetal objects, buried to a selected depth in dry sand in an indoor sandbox, are illuminated by an impulse radar system playing the role of a ground penetrating radar (GPR). The recorded time-series data of backscattered echoes from these targets are analyzed in the joint time-frequency domain using a pseudo-Wigner distribution (PWD). These distributions with their extracted features in the two-dimensional time- frequency domain are viewed as the target signatures. We have previously demonstrated the usefulness of the PWD for target identification purposes, in particular the merits of the PWD relative to various competing time-frequency distributions for targets buried at different depths. We have also used a classification method developed from the fuzzy C-means (FCM) clustering technique to reduce the number and kind of features in the PWD signature templates. This is accomplished by converting the PWD signature into a representation by a cluster of points associated with a weight and then reducing the cluster to a number of (weightless) cluster centers. We investigate here how the selected number of cluster centers and the choice of a governing parameter in the FCM algorithm influence the target recognition capability of the resulting signature representations. The classification algorithm is tested against validation data taken from an additional set of returned echoes. The same targets are used but they are buried at a different location in the sand. Class membership of a target is then decided using a simple metric. The results of our investigation serve to assess the possibility of identifying subsurface targets using a GPR, by means of the present technique.

This paper presents an algorithmic approach for the automatic detection and recognition of stationary motorized vehicles in infrared images. Covering the whole object recognition processing chain, robust solutions are proposed for the preprocessing, detection and segmentation steps, with particular emphasis on the feature extraction and final classification stages. The segmentation process consists of a graph based analysis strategy that group high level features such as hot regions and contour segments of pre-specified types into bounding rectangles of the potentially relevant objects. Feature extraction proceeds with the superimposition of a grid of a predefined number of equally sized cells onto the bounding rectangular window determined for each potential target. From the measurements evaluated for each cell it is built a feature vector that feeds a supervised neural network classifier, aiming to perform a coarse recognition of the detected targets.

In this paper, a new neural network directed Bayes decision rule is developed for target classification exploiting the target's dynamic behavior. The system consists of a feature extractor, a neural network directed conditional probability generator and a sequential Bayes classifier. The velocity and curvature sequences extracted from each track are used as the primary features. Several hidden states are used to train the neural network, the output of which is the conditional probability of occurring the hidden states given the observations. These conditional probabilities are then used as the inputs to the Bayes classifier to make the classification. The classification results are updated recursively whenever a new scan of data is received. Simulation results on both clean tracks and heavily cluttered Infrared (IR) satellite images are presented to demonstrate the effectiveness of the proposed methods.

As we reported in the last few years, a one-layered, hard- limited perceptron is generally sufficient for carrying out a robust recognition on any untrained pattern if the training class patterns satisfy a certain PLI condition. For most pattern recognition applications, this condition should be satisfied. When this condition is satisfied, an automatic feature extraction scheme can then be derived using some N- dimension Euclidean geometry theories. This automatic scheme will automatically extract the most distinguished parts of the N-vectors used in the training. These distinguished parts or the feature vectors will then allow a very robust recognition when untrained patterns are tested in the recognition mode. Theoretical derivation and live experiments revealing the physical nature of this novel, ultra-fast learning, pattern recognition system will be presented in detail.

This paper presents a method for image understanding that combines a fuzzy pixel-based feature extractor with a novel, multiple prototype classifier to detect and interpret small targets in LADAR intensity images when very few pixels on target are available. The method is based on the fuzzy c-means clustering algorithm (FCM). Prototypes are appended to an unprocessed image and low-level attributes of each pixel in the combined image are computed from the 8-neighbor pixels using FCM in 9 dimensions with 5 classes. A feature vector is then extracted from each prototype using a centered n x m window. The class membership vectors of the labeled prototypes are compared to the resulting class membership vectors of each unlabeled pixel to generate a set of confidences of a pixel's membership in the prototype classes. The fuzzy partition produced by FCM retains spatial integrity of each pixel label vector and relates the pixel level information contained in the partition to pixels in the data to be labeled. The method exhibits good behavior for images that do not contain any of the original prototype targets.

A key question in SAR-aided search is the relative utility of L-Band versus P-Band data. A continuing study is underway using target data collected by the NASA Search and Rescue Mission. This paper presents the most recent results of the investigation, including Navy P-3 SAR data at L-Band and JPL SAR data at L-band and P-band.

Search and rescue operations are often characterized by the need to search for a relatively small craft (airplane or boat), and the search may have to consider a large area when emergency location beacons have failed. The ability to locate the crash site quickly is critical to any survivors: their probability of survival decreases rapidly following the accident. Ground and airborne search and rescue operations can be hampered by inclement weather or the size or remoteness of the area to be searched. Synthetic aperture radar satellites, with the ability to image large swaths of the earth's surface through any weather condition, may offer valuable assistance. RADARSAT-2, to be launched in February, 2002, will provide users with advanced SAR imagery, having fully polarimetric modes and resolutions as fine as 3 meters. In this paper, the suitability of synthetic aperture radar satellites for support of search and rescue operations is analyzed, specifically considering the capabilities of Canada's RADARSAT-2 satellite.

Interferometric SAR exploits the coherent nature of multiple synthetic aperture radar images to recover phase (range difference) information and thence terrain evaluation data as well as other phase derivative products such as Coherent Change Detection (CCD). Of the numerous factors that can degrade the coherency of multiple SAR collections, foliage constitutes one of the most challenging. The foliage problem in IFSAR is discussed and an airborne multiple pass collection is used to illustrate some facets of the problem. Resolution as a variable in the tradeoff between the bias and variance of the interferogram is discussed in the context of the example.

One of the most promising tools for airborne SAR search and rescue is the use of interferometric techniques such as wavenumber filtering. These techniques make possible extremely accurate information extraction from SAR image pairs. The potential gain in accuracy is significant, since accuracy of measurements can theoretically be determined to within a wavelength (centimeter accuracy) as opposed to a pixel distance (meters). This paper presents the latest interferometric SAR (IFSAR) processing algorithms developed by the authors, along with examples of their use on actual data.

The Beaconless Search & Rescue Program at NASA Goddard Space Flight Center (GSFC) has been working to solve the technological challenges associated with detecting small aircraft crash sites using synthetic aperture radar (SAR) imagery. One area of work has focused on the use of fully polarimetric imagery to both improve image quality and distinguish the crash sites from the natural background. Data from aircraft based SARs have been used for development but since a SAR satellite deployment is one possible option for a practical Search and Rescue system, the work is being extended to satellite SAR imagery. This paper presents the results of processing Shuttle Imaging Radar-C (SIR-C) data collected over an aircraft crash site near Wadesboro, North Carolina through the target detection software developed at GSFC. The results demonstrate the ability to achieve crash site detection using SAR data collected from Earth orbit.

A Piper Malibu aircraft crashed on April 11, 1998 near Kalispell, Montana. After more than a month of visual searching, the official search was suspended and the missing pilot's family turned to private resources and NASA to continue the search. This paper details the NASA Goddard Space Flight Center Search and Rescue Mission's participation in the follow-on search effort.

In this paper we exploit high resolution millimeter wave radar ISAR imagery to develop a vehicle classification algorithm, which is robust to orientation and position of the vehicle in the scene. A template based approach is presented and the effect of a number of methods of creating templates investigated. To incorporate the effect of uncertainty in vehicle position and orientation, an approach based on mixture models is developed. The specification of the model is discussed and various approaches for determining the parameters of the model have been assessed. Preliminary results using mixture models to model vehicle signatures and uncertainties in position and orientation are presented. The models and techniques reported here provide a robust approach for general radar classification problems that incorporates uncertainty in a principled manner and improves generalization.

The combined effects of laser speckle and local oscillator shot noise degrade coherent laser radar range measurements. As a result, laser radar range imagery suffers from both uniformly-distributed range anomalies and Gaussian-distributed local range errors. Our goal in this research is to develop a target recognition system capable of recognizing military vehicles in range images provided by airborne laser radars. In particular, we will focus on using laser radar range imagery in a statistical model-based object recognition system. We shall present performance results for our object recognition system using laser radar data from the MIT Lincoln Laboratory Infrared Airborne Radar (IRAR) data release together with 3-D CAD models which account for the possible military targets that may be present on the site imaged by the laser radar.

Modern radar performs target recognition and target imaging tasks, in addition to conventional tasks of detection and tracking. New processing techniques, like stepped frequency wave-forms, modulation due to rotary parts, etc. and RF hardware are now becoming available and will soon result in lower-cost high resolution radar for commercial as well as military applications. Feature extraction, namely modulation due to rotary wings can be used to discriminate fixed wing verses rotary wing aircraft. Further advantage of wide band operation allows generation of synthetic range with resolution of few centimeters required for target identification. An important class of wave-forms used for high resolution mapping and target imaging, falls under the category called stretch wave-form processing. The simplest wave-form processing uses Fourier transform (FFT or IFFT). Range profiles thus generated, show the scattering centers of the target, and are being used for one-dimensional target identification procedures. These range profiles, however are very sensitive to target registration due to zero sampling inherent in the FFT procedure. This phenomenon together with the well known aspect sensitivity of the target profiles, plays havoc in the automatic target recognition procedures. In this paper we present a new method of obtaining range profiles or frequency spectrums. These spectrums do not sample zeros and are robust with respect to range motion or range registration. Based on the super-resolution techniques, analysis is given for the Rayleigh's Quotient procedure. It is shown that all the peaks of the range profiles are preserved and none of the zeros are sampled.

Automatic Target Recognition (ATR) systems are required to identify and differentiate between a large number of targets under a broad class of scenario variations. To accomplish this task ATR systems will employ high-dimensional data, such as High Range Resolution (HRR) radar data, which improves discriminability, but leads to very large databases and the attendant computational and storage requirements. Reducing the size of ATR databases without jeopardizing recognition performance is a potential solution to the above challenges. This reduction can be achieved through: (1) Feature Extraction, or (2) Vector Quantization. In this paper we apply VQ classification algorithms to measured HRR radar data to assess the effects of database compression on ATR performance. In particular, we introduce a distance-based reject option into the nearest neighbor classification scheme and perform experiments to investigate the error-reject tradeoff via error-reject curves. Experimental results indicate that a substantial compression (about 10:1) of the training database can be achieved with little degradation of ATR performance on the measured HRR database.

The quality of radar target recognition can be enhanced by exploiting its polarization signatures. A specialized X-band polarimetric radar was used for target recognition in experimental investigations. The following polarization characteristics connected to the object geometrical properties were investigated: the amplitudes of the polarization matrix elements; an anisotropy coefficient; depolarization coefficient; asymmetry coefficient; the energy of a backscattering signal; object shape factor. A large quantity of polarimetric radar data was measured and processed to form a database of different object and different weather conditions. The histograms of polarization signatures were approximated by a Nakagami distribution, then used for real- time target recognition. The Neyman-Pearson criterion was used for the target detection, and the criterion of the maximum of a posterior probability was used for recognition problem. Some results of experimental verification of pattern recognition and detection of objects with different electrophysical and geometrical characteristics urban in clutter are presented in this paper.

Traditionally, target tracking has been used for aerospace applications, such as, tracking highly maneuvering targets in a cluttered environment for missile-to-target intercept scenarios. Although the speed and maneuvering capability of current aerospace targets demand more efficient algorithms, many complex techniques have already been proposed in the literature, which primarily cover the defense applications of tracking methods. On the other hand, the rapid growth of Global Communication Systems, Global Information Systems (GIS), and Global Positioning Systems (GPS) is creating new and more diverse challenges for multi-target tracking applications. Mobile communication and computing can very well appreciate a huge market for Cellular Communication and Tracking Devices (CCTD), which will be tracking networked devices at the cellular level. The objective of this paper is to introduce a new concept, i.e., Virtual Target Tracking (VTT) for commercial applications of multi-target tracking algorithms and techniques as applied to mobile satellite communication networks. It would be discussed how Virtual Target Tracking would bring more diversity to target tracking research.

In this paper, a novel video indexing scheme is proposed for moving target in videos acquired from active visual systems. By virtue of the object-centered nature of active videos, the extremely difficult correspondence problem is simplified greatly. With the Foveate Wavelet Transform, a novel variable resolution technique with desirable emulation of the animate vision system and direct image processing, as the representation of active videos, the object description for the object of primary interest during smooth pursuit can be generated in the foveal area based on the behavior of the motion field estimated from the two-pass MRF-MAP algorithm for motion estimation based on Mean Field theory. Afterwards, an illumination-invariant compressed histogram, merely 36 numbers, is generated as the index for this object.

The problem of estimating aircraft pose information from mono- ocular image data is considered using two different pose estimation algorithms. Both algorithms are based on the rotation invariant moment approach that was introduced by Dudani. The dependence of pose estimation accuracy on image resolution and aspect angle was investigated through simulations using sets of synthetic aircraft images. It is shown that increased pose-estimation accuracy can be obtained by breaking the nearest neighbor search procedure in two parts.

In this paper, according to the translation- and scale- invariant adaptive wavelet decomposition we proposed before, we define a multiresolution energy distribution signature (MES) of textures, which is stable under translation, scale and gray scale transforms, and more, by using MES, we present a method for scale-invariant texture pattern recognition. At end, we give experiments to demonstrate the validity of our scheme.

Third Generation Automatic Target Acquisition (ATA) technology must meet three important design criteria to be viable for 21^st Century Missile Applications: (1) it must be robust to sensor and signature variations, (2) it must solve the geometric distortion problems resulting from different reconnaissance and missile sensor perspectives, and (3) it must be extremely fast, extremely intuitive, and extremely flexible to mission plane. This paper describes an approach under development at Raytheon Systems Company, which meets these requirements for a large class of targets. The approach uses passive reconstruction of key scene features in real-time to produce a 3D scene description. This description can then be manipulated to correspond to the mission planning perspective and matched in that domain. The acquisition algorithm seamlessly transitions to track once a lock on has been established. We will describe our approach from mission planning to 3D terminal aimpoint placement and demonstrate algorithm performance in both closed-loop synthetic and open- loop captive flight test environments. We will also discuss extensions of this algorithmic approach to other sensor types (primarily LADAR) and additional target sets.

Tanner Research, Inc. has developed an arbitrary-scale wavelet constant false alarm rate detector (CFAR) for automatic target recognition (ATR). The proposed effort will be the first exploration of the connection between a traditional CFAR detector and a wavelet transient detector. Traditional CFAR detectors use a fixed, pre-selected kernel scale. Our detector will instead automatically select the scale most beneficial scale for image analysis. Our optimal-scale wavelet CFAR reduces the need for computationally expensive classifiers; thus, it improves image analysis efficiency.

The use of polarization-sensitive sensors is being explored in a variety of applications. Polarization diversity has been shown to improve in a significant way the performance of the automatic target detection and recognition. However, it also brings out the problems associated with processing and storing more data and the problem of calibration of the polarimetric sensors themselves and the problem of dealing with polarimetric sensors on board of maneuvering platforms. In this paper, we present techniques for extracting attributes that are invariant to the polarization transformation. The polarimetric target signatures are represented in terms of the components of the Jones vectors. The cases associated with active radar and passive infrared polarimetry are considered. For the radar, two algebraic expressions, functions of the radar cross sections, exist that are invariant under any linear transformation of the basis of the polarization. In the infrared case, invariant algebra is used to extract the target-related attributes that are invariant to the affine transformation of the polarization coordinate system.

This paper describes a biologically motivated visual architecture for automatic target acquisition and tracking. The model, that is based on principal characteristics of the Human visual System (HVS), was incorporated into a prototype ATR testbed that performs multi-resolution target signature extraction at the sensor level. The extracted target features are then integrated into a consistent representation of the scene using a parallel attention model of the HVS. The described ATR solution integrates a number of innovations on target segmentation, camouflage elimination, 3D invariant target identification and intelligent tracking into a concise framework. The architecture is transparent to sensor technology. Simulation and experimental results are presented.

Despite much prior work, one of the problems that still persists in using composite correlation filters for Automatic Target Recognition (ATR) is the high false alarm rate due to clutter. In this paper, we propose two methods (one based on clustering and another based on extending the maximum average correlation height or MACH filter) to improve the clutter rejection capability of composite filters. Initial numerical results are presented to illustrate the potential improvements.

Synthetic Discriminant Function (SDF) filters are characterized by hard constraints placed on correlation peak values. It is shown that we can obtain more control on the clutter rejection performance of the SDF filters by using complex constraints. Also, analytical expressions are derived that connect the average correlation peak intensity to constraint phases.

The impact of wavelet based compression on automatic target recognition (ATR) is investigated by applying wavelet compression to test scenes. The correlation algorithms known as maximum average correlation height (MACH) filter and the distance classifier correlation (DCCF) filter are used for ATR. The impact of compressing the correlation filters is also studied. The wavelet compression algorithm makes use of a progressive technique of embedded zerotree wavelet coding followed by adaptive arithmetic coding. Two target data sets are used for testing and training in this study. The first is composed of infrared (IR) images of a T72 tank and BMP armored personnel carrier. The second is a set of synthetic aperture radar (SAR) targets from the publicly released Moving and Stationary Target Acquisition and Recognition (MSTAR) database.

Correlation filters have been successfully utilized for object detection in many applications. Each sensor type, however, presents different advantages and challenges. This paper describes the application of correlation filter techniques for automatic target recognition (ATR) to Laser Radar (LADAR) sensor images. Filters are designed using synthetic models, and incorporate range and aspect tolerance for mobile objects. The model generator easily takes into account the sensor field of view (FOV), look-down angle, ground cell size, and shadows. The filters are also designed to exploit various coordinate transforms that are feasible with a LADAR sensor. The correlation algorithm has the unique potential of exploiting the intensity information in conjunction with the range measurements provided by the LADAR. Examples using fixed and mobile targets are presented, along with statistical performance results.

In this paper, we incorporate the theory of the multidimensional discriminant analysis into a photorefractive correlator architecture. In this approach, a set of eigenimages extracted from a large number of training images by K-L transform are stored in a photorefractive crystal by using the two-wave mixing volume holographic storage technique and used as the reference images in the photorefractive correlator. When any new image inputs the correlator, angularly separated beams with different light intensities are obtained simultaneously. They represent the optical correlation results between the input and the set of eigenimages and can be regarded as eigenfeatures. Then the multidimensional discriminant analysis will be applied to these features for training and classification. During both processes, a bifurcating tree structure is used, by which the recognition speed of the system can be greatly improved. This approach takes the advantages of both the high degree of parallelism of the photorefractive correlator and the optimal discriminating ability of the multivariate statistical methods for classification.

In many image analysis problems it is possible to take advantage of the structural relationships between various parts of the objects being imaged in order to index the images of the objects. For example, many satellites consists of a main body and outlying sub-components. Thus, in many circumstances satellites can be indexed in a model database by the distinct structural relationships between their sub- components. However, algorithms based on structured sub- components necessitate the use of robust and reliable 2-D image segmentation techniques to successfully partition images into their sub-components. Unfortunately, this segmentation task can be highly problematic for objects with complex components and under harsh, unfavorable lighting conditions. The research presented here describes a new method to compute indices which can be used for image indexing without image segmentation. We use satellite imagery as a convenient image class for which to demonstrate our method. Our method partitions the image into many small equal-area pieces. We refer to this technique as differentiation. Differentiated images result in a set of sub-images that collectively represent the structural information inherent in the image. We prove that a primitive matrix with at most four non-zero eigenvalues can be constructed from the differentiated image. This property (1) significantly reduces storage requirements for a model database, (2) reduces the computational burden of subsequent recognition processes, and (3) supports an efficient and accurate matching procedure. To evaluate the efficiency of our algorithm for a recognition application, we use boundary methods as a feature set evaluation method to quantify the utility of the eigen-indexes obtained by our method as compared to other existing indexing methods.

The motive for this study is to improve learning based target recognition by utilizing discriminative learning for minimum error classification and applying discriminative feature extraction based on a preset criterion. Radar cross section measurements of four commercial aircraft, obtained experimentally in a compact range, are used for training and testing a three layered back propagation neural network for target identification purposes. It is assumed that the aspect angles (or azimuth positions) of all four targets are either completely known or known within 20 degrees uncertainty range. The scattering parameters of each target are assigned selective weights and presented to a discriminative feature extractor. The performance of the proposed target recognition system is examined assuming different noise scenarios and various levels of azimuth ambiguity. The proposed scheme is also tested in scenarios where the maximum likelihood approach is available and the performances of both recognition techniques are compared. Issues concerning the number of hidden nodes, training parameters, and weight convergence are discussed.

We present a fundamental approach to the processing of signals for the detection of targets immersed in clutter in any type of digitized image (synthetic aperture radar, optical, acoustical, etc.), analogous to the standard theory of target detection for pulsed radars, with applications to aided target recognition (ATR). Expanding upon recent results of DiPietro and Fante, we derive a new ATR detection probability function analogous to the Swerling Type-4 detection probability function of pulsed-radar detection theory. We carry out a comparative theoretical and numerical analysis of the single- look and single-pulse probabilities of detection of targets in the general ATR and pulsed-radar cases, and also provide a comparative analysis of the noncoherent integration of multiple samples of image data in the ATR and pulsed-radar cases. We derive expressions for the binary integration of all single-look ATR detection probabilities, and perform a comparative theoretical and numerical analysis of the performance characteristics of binary integration versus noncoherent integration. A detailed numerical analysis of the optimization of parameters for peak ATR binary integration performance in low-resolution and/or low signal-to-clutter ratio images is performed, and a comparative analysis of optimization of parameters for the pulsed-radar theory in low signal-to-noise ratio environments is carried out.

For the exploitation of aerial and satellite imagery, human military photo interpreters need support by automatic image analysis components to meet the requirements of large data set analysis under strong time constraints. Extending the approaches of performance analysis of automatic target detection, a concept and an experimental study for the assessment of machine assisted vehicle detection is presented. This evaluation pursues the following goals: Extraction of a usability measure in terms of algorithm performance combined with user-oriented parameters. Secondly, an extraction of requirements for the image exploitation process concerning the algorithm performance, the man-machine interface and the training of the photo interpreters. A performance analysis concept for vehicle detection algorithms is presented as well as an experimental setup of the whole interactive exploitation process. This setup has been applied in an experiment with more than 100 real images and more than 40 military photo interpreters.

This contribution presents a comprehensive framework for algorithm evaluation. When we speak of evaluation, we have in mind that first the performance of an algorithm is measured and then the measured performance is assessed with regard to a given application. The performance assessment is done by applying an assessment function that uses desired values for the performance measures and weighting factors giving the importance of each measure, thus considering the application- specific requirements. The algorithm evaluation's goal is to verify the specification of an algorithm. This specification is mainly given by the definition of the input data and the expected output data, both of which are determined by the application. Prior to the evaluation process the algorithm specification has to be laid down by analyzing the application in order to deduce its requirements as well as by defining the application relevant data sets. To organize this sequence of preparatory steps and to formalize the accomplishment of the evaluation we have developed a 3-phase approach, consisting of the definition phase, the tuning phase, and the evaluation phase. An extensive software toolbox has been developed to support the evaluation process.

This paper presents a three-dimensional (3-D) receiver operating characteristic (ROC) trajectory which combines the principles of the standard ROC curve used in automatic target recognition (ATR) with a third performance measure. These 3-D ROC trajectories are used to compare competing target recognition algorithms when unknown targets are present in the data. A probability of rejection is added to the standard ROC curve that includes the probability of successful detection (true positive rate) and the probability of false alarm (false positive rate). By using the probability of rejection, targets may be eliminated that are difficult to classify, or perhaps unknown. The 3-D ROC trajectory can be a useful tool for a SAR image analyst for understanding the tradeoffs between the probability of rejection and the two standard performance measures commonly used in detection problems.

A wave-based matching-pursuits algorithm is used to parse multi-aspect time-domain backscattering data into its underlying wavefront-resonance constituents, or features. Consequently, the N multi-aspect waveforms under test are mapped into N feature vectors, y_n. Target identification is effected by fusing these N vectors in a maximum-likelihood sense, which we show, under reasonable assumptions, can be implemented via a hidden Markov model (HMM). In this paper, we utilize a continuous-HMM paradigm, and compare its performance to its discrete counterpart. Algorithm performance is assessed by considering measured acoustic scattering data from five similar submerged elastic targets.

A new method for spectral unmixing is developed. This method accounts for the nonnegativity of the mixing constants. Preliminary experiments to test the method are reported.

We report results from a demonstration of a MWIR non-scanning, high speed imaging spectrometer capable of simultaneously recording spatial and spectral data from a rapidly varying target scene. High speed spectral imaging was demonstrated by collecting spectral and spatial snapshots of filtered blackbodies and combustion products. The instrument is based on computed tomography concepts and operates in a mid-wave infrared band of 3.0 to 4.6 micrometer. Raw images were recorded at a video frame rate of 30 fps using a 160 X 120 InSb focal plane array. A reconstruction of a simple object is presented.

Recent advances in the field of spectral sensing technology have elucidated the benefits of multi-spectral and hyperspectral sensing to the Army's user community. These advancements, when properly exploited can provide the Army with additional and improved automated terrain analysis, image understanding, object detection, and material characterization capabilities. The U.S. Army, led by the Topographic Engineering Center, has established a Center of Excellence for Spectral Sensing Technology. This Center conducts Army wide collaborative research on, and development and demonstration of spectral sensing, processing and exploitation technologies. The Center's collaborative efforts integrate Army programs across multiple disciplines and form a baseline program consisting of coordinated technology thrusts. The program's applied research and demonstration components will in turn support an Army spectral Strategic Technology Objective (STO) that will ultimately support and leverage joint service efforts starting in FY00. Existing efforts span the domains of sensor hardware, data processing architectures, algorithms, and, signal processing and exploitation technologies across wide spectral regions. These thrusts in turn enable progress and performance improvement in the automated analysis, understanding, classification, discrimination, and identification of terrestrial objects, and materials. The participants draw upon common scientific processes and disciplines to attack similar problems related to different categories and domains of phenomenology. This paper describes the Center's program and objectives along with an explanation of the Army's strategy and approach in support of its program objectives.

For an on-board automatic target recognition (ATR) system to be useful to the crew of a military platform, the ATR must reduce the mission risk or increase its lethality. This utility may be increased by shortening the operator's time to interrogate possible threat targets or by enabling weapon deployment at a greater range. Obstacles to deployment of ATRs have included an excess of false cues and difficulty in adapting developmental configurations to processing architectures that can operate in the required environmental conditions without serious performance degradation. We present a real-time FLIR ATR software architecture that is scaleable across multiple processors and readily portable to a number of hardware platforms. Fusion with cues from an on- or off-board synthetic aperture radar (SAR) provides a significant reduction in the amount of processing required to classify targets while simultaneously increasing the confidence in each target hypothesis. The FLIR ATR and fusion are implemented on commercial off-the-shelf (COTS) processors that are available in ruggedized versions, and the software is constructed to allow portability to other processor families without major disturbance to those parts of the code that embody the algorithm content.

Hyperspectral imaging holds great promise for object detection and recognition due to the richness of the spectral content in images from such objects. Ordinary broad-band imagers integrate the spectral information over the entire spectral band of coverage when used without any spectral filtering. In general, the spectral details in the images can be obtained by using an optical filtering element such as a filter wheel, a grating, or an acousto-optic tunable filter (AOTF). Since each task of detection and object recognition may require only a limited set of specific spectral bands based on the object as well as the background, it is best to choose a filtering optical element that has high-speed spectral selectivity with high resolution. Of all the optical filtering elements available, only an AOTF offers this capability. Such capability greatly reduces the amount of data collection and processing. In this paper, we present hyperspectral images obtained in the laboratory and from field tests, using visible-to-near-IR (VNIR) AOTF imagers. The imagers use a tellurium dioxide, TeO₂, AOTF cell that covers the spectral band from 450 to 1000 nm with a spectral resolution of 10 nm at 600 nm, a charged coupled device (CCD) camera, image-forming optics, frame grabber board, rf electronics, and control and processing software. The imager used for outdoor testing is equipped with a variable phase retardation plate to obtain images with polarimetric signatures (patent pending). The spectral and polarimetric imaging capabilities of the AOTF imager were successfully tested to discriminate targets and backgrounds in various environments.

This paper describes an automatic target detection algorithm based on the sequential multi-stage approach. Each stage of the algorithm uses more spectral bands than the previous stage. To ensure high probability of detection and low false alarm rate, Chebyshev's inequality test is applied. The sequential approach enables a significant reduction in computational time of a hyperspectral detection system. The Forest Radiance I database collected with the HYDICE hyperspectral sensor at the U.S. Army Proving Ground in Aberdeen, Maryland is utilized. Scenarios include targets in the open, with footprints of 1 m and different times of day. The total area coverage and the number of targets used in this evaluation are approximately 6 km² and 126, respectively.

An iterative algorithm for fitting a simplex around a set of data vectors is proposed. The algorithm is a gradient descent on an objective function defined on the vertices. This objective function is the sum of two terms: the first term is the volume of the simplex, which we would like to minimize, and the second term is a penalty term which has the effect of 'pushing' the faces of the simplex away from the data points. The gradient of each of these terms is determined analytically, and used in the gradient descent algorithm. The penalty term includes a multiplicative constant which approaches zero as the gradient descent algorithm progresses, causing the iterates to converge to the vertices of a simplex which fits tightly around the given data points. The performance of this algorithm is demonstrated using simulated data generated from real spectral libraries.

In this paper we empirically examine the probability of detection (P_D) and false alarm rate (FAR) for crash site detection using polarimetry to discriminate between aircraft target signatures within natural clutter. To date, the search and rescue program has tried several automatic target recognition (ATR) algorithms from the literature. While P_D seems reasonable with these algorithms, the FAR is too high (approximately 10's per square kilometer). The objective of our analysis is to determine if this is a limitation of the ATR algorithms tried, or if this is the best that can be hoped for given the polarimetric statistics of the target and clutter.

Abstract not available.

Abstract not available.