Search

This paper presents a new approach of improving automatic target recognition (ATR) performance by tuning adaptively the Gabor filter. The Gabor filter adopts the network structure of two layers, and its input layer constitutes the adaptive nonlinear feature extraction part, whereas the weights between output layer and input layer constitute the linear classifier. From the statistic property of high-resolution range profile (HRRP), its extracted nonstationarity degree of features is tracked to extract the discriminative features of Gabor atoms. Two experimental examples show that the Gabor filter approach with simple structure has higher recognition rate in radar target recognition from HRRP as compared with several existing methods.

We propose a novel automatic target recognition (ATR) system for classification of three types of ground vehicles in the moving and stationary target acquisition and recognition (MSTAR) public release database. First, MSTAR image chips are represented as fine and raw feature vectors, where raw features compensate for the target pose estimation error that corrupts fine image features. Then, the chips are classified by using the adaptive boosting (AdaBoost) algorithm with the radial basis function (RBF) network as the base learner. Since the RBF network is a binary classifier, we decompose our multiclass problem into a set of binary ones through the error-correcting output codes (ECOC) method, specifying a dictionary of code words for the set of three possible classes. AdaBoost combines the classification results of the RBF network for each binary problem into a code word, which is then 'decoded' as one of the code words (i.e., ground-vehicle classes) in the specified dictionary. Along with classification, within the AdaBoost framework, we also conduct efficient fusion of the fine and raw image-feature vectors. The results of large-scale experiments demonstrate that our ATR scheme outperforms the state-of-the-art systems reported in the literature.

Due to recent advances in hyperspectral imaging sensors many subtle unknown signal sources that cannot be resolved by multispectral sensors can be now uncovered for target detection, discrimination, and identification. Because the information about such sources is generally not available, automatic target recognition (ATR) presents a great challenge to hyperspectral image analysts. Many approaches developed for ATR are based on second-order statistics in the past years. This paper investigates ATR techniques using high order statistics. For ATR in hyperspectral imagery, most interesting targets usually occur with low probabilities and small population and they generally cannot be described by second-order statistics. Under such circumstances, using high-order statistics to perform target detection have been shown by experiments in this paper to be more effective than using second order statistics. In order to further address a challenging issue in determining the number of signal sources needed to be detected, a recently developed concept of virtual dimensionality (VD) is used to estimate this number. The experiments demonstrate that using high-order statistics-based techniques in conjunction with the VD to perform ATR are indeed very effective.

Clutter plays a very important role in the area of human-in-the-loop target acquisition. A great deal of interest has recently been shown in defining clutter for electro-optical images. In spite of so many definitions available, no single clutter definition has been agreed on by the electro-optical community as being the best. We modify the structure similarity index, which is extensively used for image quality assessment, to be clutter metrics and call them target structure similarity measures (TSSIMs). Two different methods for calculating the TSSIM metrics are suggested. The TSSIMs of many images in Search_2 dataset have been analyzed. Experimental results show the proposed metrics correlate very well to the detection probability of the observers.

A technique for target location estimation in through-the-wall radar imaging applications is presented. The algorithm corrects for the shifts in target positions due to ambiguities in the wall thickness and dielectric constant. We consider uniform walls and perform imaging using wideband beamforming, with the antennas placed against the wall. Behind-the-wall images are obtained using different structures of transmit and receive arrays. For each array structure, a trajectory of the shifts in the target locations is generated assuming different wall parameters. The target position is estimated as the intersection of the corresponding trajectories. The paper shows that for unknown wall thickness or dielectric constant, the point of intersection is the true target position. In the case when both parameters are unknown, the estimated target location is in close proximity to the target true position. It is demonstrated that the performance of the proposed technique is rather insensitive to the target location behind the wall and to various array structures.

An automatic target recognition (ATR) algorithm, based on greedy learning of Gaussian mixture model (GMM) is developed. The GMMs were obtained for a wide range of ground surveillance radar targets such as walking person(s), tracked or wheeled vehicles, animals, and clutter. Maximum-likelihood (ML) and majority-voting decision schemes were applied to these models for target classification. The corresponding classifiers were trained and tested using distinct databases of target echoes, recorded by ground surveillance radar. ML and majority-voting classifiers obtained classification rates of 88% and 96%, correspondingly. Both classifiers outperform trained human operators.

A new approach to the detection of masked targets and the estimation of target parameters with significant computational savings over standard approaches to this problem based on the CLEAN algorithm is proposed. The approach implements the CLEAN process efficiently as only few selected samples are used rather than using all samples from the matched filter output.

In this paper the track initiation problem is formulated as multiple composite hypothesis testing using maximum likelihood estimation with probabilistic data association (ML-PDA). The hypothesis selection is based on the minimum description length (MDL) criterion. We first review some well-known approaches for statistical model selection and the advantage of the MDL criterion. Then we present one-dimensional examples to illustrate the MDL criterion used in multiple composite hypothesis testing and the performance limit of the ML-PDA for track initiation is interpreted in terms of the sharpness of the hypothesis testing. Finally, we apply the MDL approach for the detection and initiation of tracks of incoming tactical ballistic missiles in the exo-atmospheric phase using a surface-based electronically scanned array (ESA) radar. The targets are characterized by low signal-to-noise ratio (SNR), which leads to low detection probability and high false alarm rate. The target acquisition problem is formulated using a batch of radar scans to detect the presence of up to two targets. The ML-PDA estimator is used to initiate the tracks assuming the target trajectories follow a deterministic state propagation. The approximate MDL criterion is used to determine the number of valid tracks in a surveillance region. The detector and estimator are shown to be effective even at 4.4 dB average SNR.

In military affairs, we often have to make decisions in conditions of uncertainty when the available situation, enemy and friendly force data are incomplete or unreliable. One of the solutions [...]

Seeking probability models for images, we employ a spectral approach where the images are decomposed using bandpass filters and probability models are imposed on the filter outputs (also called spectral components). We employ a (two-parameter) family of probability densities, introduced in [11] and called Bessel K forms, for modeling the marginal densities of the spectral components, and demonstrate their fit to the observed histograms for video, infrared, and range images. Motivated by object-based models for image analysis, a relationship between the Bessel parameters and the imaged objects is established. Using [L.sup.2]-metric on the set of Bessel K forms, we propose a pseudometric on the image space for quantifying image similarities/differences. Some applications, including clutter classification and pruning of hypotheses for target recognition, are presented. Index Terms--Image probabilities, spectral analysis, Bessel K forms, clutter classification, target recognition, Gabor filters.

Some views of Professor James Paul Helferty regarding the use of turn-rate distributions for acceleration modeling invite comments. The equations sigma(super 2, sub x) and sigma(super 2, sub y) should have their derivations with variances r(sub x) and r(sub y) as sigma(super 2, sub x) equals sigma(super 2, sub r) cosine(super 2)-beta plus sigma(super 2, sub beta) times r(super 2) times sine (super 2)-beta and sigma(super 2, sub y) equals sigma(super 2, sub r)times sine(super 2)-beta plus sigma(super 2, sub beta) times r(super 2) times cosine(super 2)-beta.

Automatic target recognition (ATR) is a multidisciplinary field that involves computer processing of data from numerous sensors for aided or autonomous target detection and recognition. A variety of techniques can be used for target detection and ATR in combination with other related purposes. Multiresolution processing for clutter modeling, physics-based processing for change detection, and model-based image sequence processing for change and motion detection are some of the ATR techniques. ATR requires diverse technology and expertise, has wide applications, and a challenging future.

The performance analysis of source trajectory estimation using measurements provided by multiple platforms (or arrays) is studied. In numerous practical situations, the maneuvering ability of the receiver (e.g., a ship towing linear arrays) is limited, leading to the assumption that the observer motion is rectilinear and uniform. Even if this hypothesis appears quite restrictive, practical and tactical considerations fully justify its interest This leads to consider multiple (platform) target motion analysis (denoted MTMA) and to analyze the performance of such trajectory estimation methods.

We describe a correlation-based distance-classifier scheme for the recognition and the classification of multiple classes. The underlying theory uses shift-invariant filters to compute distances between the input image and ideal references under an optimum transformation. The original distance-classifier correlation filter was developed for a two-class problem. We introduce a distance-classifier correlation filter that simultaneously considers multiple classes, and we show that the earlier two-class formulation is a special case of the classifier presented. Initial results are presented to demonstrate the discrimination- and distortion-tolerance capabilities of the proposed filter. Key words: Correlators, correlation filters, target recognition, distance classifier, shift invariance, distortion tolerance, discrimination, Fourier transforms.

The state-observability problem for passive target tracking by angle and/or frequency measurements is analyzed. The method used is independent of the target model and can be specialized to arbitrary models in a systematic way. As an example, the general Nth-order dynamics target model has been discussed. In this way the previously established necessary and sufficient observability conditions are rederived in a much simpler way and extended to more general cases.

This work focuses on the use of actual radar sensor data to construct a multifeature decision space formulation to the target identification (TID) problem. The decision space concepts are classical and the basic target features used can be extracted from the data provided by standard radar sensor operating modes. The target features used in the decision process along with the construction of a useful statistical description of each feature for a given target are presented. A multidimensional formulation of the decision space and the decision logic is also presented leading to a versatile multifeature based algorithm. The algorithm performance has been evaluated on live data and the results are reported.

A hybrid estimator is formulated by combining the estimates from an empirical Bayes estimator and a jackknifed Bayes estimator. This hybrid estimator performs better than the maximum likelihood estimator, when used to estimate the cumulative probability of detection of surveillance radars.

An explicit, closed-form formulae of advanced guidance laws for a linear, time-invariant, acceleration-constrained arbitrary-order missile, and a linear, time-invariant, arbitrary-order, randomly maneuvering target with noisy position measurements are derived. Two approaches are presented. The first approach derives the optimal guidance law for a quadratic objective. The solution is the guidance law for deterministic system with limiting on the commanded acceleration applied on the estimated state. The limiting function in this case is the saturation function. The second approach derives a control law called the average input guidance law. This approach is based on the idea of applying the average of the input that would have been applied to the plant if the noises were known. The solution has similar structure. It is the guidance law for deterministic system with limiting on the commanded acceleration applied on the estimated state. The limiting function in this case is from the family of describing functions of the saturation function. The formulas of the different guidance laws are given in terms of the transfer function of the missile and acceleration constraint, the shaping filter of the maneuver of the target, responses to initial conditions, error variance matrix of the estimated state and weights in the criterion. It is demonstrated by simulations that although the optimal guidance law has improved performance in terms of the miss distance, the suboptimal average input guidance law consumes less energy.

A parameter, the matching score, is used to discuss the correlation property between range profiles of the same target or different targets. The property that the matching score between range profiles of two consecutive aspects from the same target is greater than that from different targets is utilized to correlate tracks in the multiple-target tracking problem. A data association algorithm based on comparing matching scores is proposed. How the aspect variation size and signal to noise ratio affect the success of data association is discussed.

A tasked tracking system updates target tracks by aiming its sensors at the expected target locations, and searching the neighborhoods of these locations until the targets are redetected. A renewal model for the state of a tracking system results in coupled partial differential and integral equations for the density of the 'time since last detection process.' One can use the densities to correlate tracking capacity with sensor resources. Exact solutions and approximate methods are discussed.

Observability is analyzed for the target tracking with bearings-only measurements (BOM) applied to a passive homing missile system using augmented proportional navigation guidance (APNG). Based on the analysis, a practical homing guidance law is proposed to enhance observability for tactical missile applications. The new guidance law is tailored to provide oscillation of line-of-sight (LOS) angle induced by initial missile heading errors without sacrificing terminal effectiveness. Simulation studies indicate that the proposed guidance law provides convergence of the state estimates essential to homing guidance applications as well as terminal effectiveness.

When the problem of tracking maneuvering targets with multiple sensors is considered in the literature, the number and type of sensors that support a given target track is usually fixed with respect to a given target location. However, in many multisensor systems, the number and type of sensors supporting a particular target track can vary with time due to the mobility, type, and resource limitations of the individual sensors. This variability in the configuration of the sensor system poses a significant problem when tracking maneuvering targets because of the uncertainly in the target motion model. A Kalman filter is often employed to filter the position measurements for estimating the position, velocity, and acceleration of a target. When designing the Kalman filter, the process noise (acceleration) variance [Q.sub.k] is selected such that the 65 to 95% probability region contains the maximum acceleration level of the target. However, when targets maneuver, the acceleration changes in a deterministic manner. Thus, the white noise assumption associated with the process noise is violated and the filter develops a bias in the state estimates during maneuvers. If a larger Q is chosen, the bias in the state estimates is less during a maneuver, but then Q poorly characterizes the target motion when the target is not maneuvering and the filter performance is far from optimal. Here, the problem of tracking maneuvering targets with multiple sensors is illustrated through an example involving target motion in a single coordinate in which it is shown that with two sensors one can have (under certain conditions that include perfect alignment of the sensors) worse track performance than a single sensor. The Interacting Multiple Model (IMM) algorithm is applied to the illustrative example to demonstrate a potential solution to this problem of track filter performance.

The idea of adjusting the detection thresholds adaptively to enhance the performance of an overall tracking system has been one of the important areas studied in tracking community for the last ten years. However, most of the previous work was developed for single target environments where a simple algorithm such as nearest neighbor (NN) or Probabilistic Data Association (PDA) filter was assumed to be used in the tracking system. We study the issues of adaptive detection thresholds based on the assumption that an optimal assignment algorithm is adopted for a multitarget and cluttered environment. This research is motivated by an important earlier work which makes the analytical evaluation of the optimal assignment algorithm possible. The performance measures considered for determining detection thresholds are the correct association probability and the expected estimation error. The analytical results obtained here represent the upper bound of the tracking performance and can be used for designing and evaluating a tracking system.

The performance of distributed constant false alarm rate (CFAR) detection with data fusion both in homogeneous and nonhomogeneous Gaussian backgrounds is analyzed. The ordered statistics (OS) CFAR detectors are employed as local detectors. With a Swerling type I target model, in the homogeneous background, the global probability of detection for a given fixed global probability of false alarm is maximized by optimizing both the threshold multipliers and the order numbers of the local OS-CFAR detectors. In the nonhomogeneous background with multiple targets or clutter edges, the performance of the detection system is analyzed and its performance is compared with the performance of the distributed cell-averaging (CA) CFAR detection system.

The problem of target motion analysis (TMA) has been the subject of an important literature. However, present methods use data estimated by a short time analysis (azimuths, Dopplers, etc.). For far sources, the nonstationarities of the array processing outputs, induced by the sources motion may be simply modeled. This model leads to consider directly a spatio-temporal TMA. Then new (spatio-temporal) data can be estimated. These estimates correspond to a long time analysis. Further, note that they are estimated independently of the (classical) bearings. In this general framework, the concept of source trajectory replaces the classical instantaneous bearings. Corresponding TMA algorithms are then studied. Then the study of statistical performance is carefully studied.

In numerous practical situations a large array of sensors is required, especially for improving the system ability to detect and track moving sources. However, the array performance is strongly affected by the lack of stationarity of the impinging signals. The use of a large array of sensors allow us to perform the source localization by using the wavefront curvature. The analysis of the range estimates based on the wavefront curvature have been thoroughly performed by various authors but, to our knowledge, is restricted to the case of fixed sources. The study of the effects of source motion for range estimation (Section III) and source detection (Section V) constitutes an important part of this work. This study reveals that the array performance can be far from the expected values, especially for large array. Therefore, the robustness of a direct processing of the whole sensor array is very questionable. The computational burden constitutes another serious drawback of this (direct) approach. The use of partitioned processings (Sections IV and VI) seems thus quite promising even if (and perhaps because) it is not optimal so as to fulfil the computation cost and robustness requirements. The performance of such processings for estimating the range of moving sources is presented.

Multihypothesis tracking (MHT) methods, well known for conventional radar data-processing, are generalized for phased-array radar application. We consider single targets in a cluttered environment. Suitable revisit and beam positioning algorithms are derived and optimized regarding the energy spent for tracking. In particular the destabilization of tracks due to clutter interference is analyzed. A comparison with a conventional algorithm (probabilistic data association filter (PDAF)) shows the saving of radar load which can be achieved by the proposed method.

An algorithm is presented to initiate tracks of a ballistic missile in the initial exoatmospheric phase, using line of sight (LOS) measurements from one or more moving platforms (typically satellites). The major feature of this problem is the poor target motion observability which results in a very ill-conditioned estimation problem. The Gauss-Newton iterative least squares minimization algorithm for estimating the state of a nonlinear deterministic system with nonlinear noisy measurements has been previously applied to the problem of angles-only orbit determination using more than three observations. A major shortcoming of this approach is that convergence of the algorithm depends strongly on the initial guess. By using the more sophisticated Levenberg-Marquardt method in place of the simpler Gauss-Newton algorithm and by developing robust new methods for obtaining the initial guess in both single and multiple satellite scenarios, the above mentioned difficulties have been overcome. In addition, an expression for the Cramer-Rao lower bound (CRLB) on the error covariance matrix of the estimate is derived. We also incorporate additional partial information as an extra pseudomeasurement and determine a modified maximum likelihood (ML) estimate of the target state and the associated bound on the covariance matrix. In most practical situations, probabilistic models of the target altitude and/or speed at the initial point constitute the most useful additional information. Monte Carlo simulation studies on some typical scenarios were performed, and the results indicate that the estimation errors are commensurate with the theoretical lower bounds, thus illustrating that the proposed estimators are efficient.

Most radar targets are complex objects and produce a wide variety of reflections. Different targets often require different models to characterize the varied statistical nature of these responses. Several commonly used statistical models and extensions to them are considered here along with techniques for' reliably calculating the resulting probabilities of detection. Marcum [1] and Swerling [2-4] have provided us with expressions for constant and some fluctuating targets, respectively. Expressions for the probability of detection for cell-averaging constant false alarm rate (CFAR) have been given by Mitchell and Walker [8], but only for clutter cells whose input to the receiver was assumed to be zero mean Gaussian clutter. These CFAR results have been generalized here for non-zero mean clutter. We have replaced the chi-square distribution for clutter with a noncentral chi-square distribution allowing for a non-zero mean. Other fluctuation models, such as log-normal, Weibull, and Ricean models, are considered for both target and clutter fluctuations. Methods for false alarm calculations are also given. Reliable algorithms for the computations of these detection probabilities are determined for the various models. Though reliability is primary, efficiency, generality and accuracy are also issues of concern.

Experiments to study the impact of signal injection and performance feedback for enhanced detection of sonar signals revealed that it had a positive effect. The experiment, involving trained sonar operators and prototype displays of an advanced sonar system, showed a significant improvement in target detection performance. The deployment of advanced sonar system caused the sonar operators to spend a substantial part of their time observing the search display.

We obtain the probability of target detection using a multifrequency, noncoherently integrated pulse burst under the following conditions. 1) Each N pulse burst contains M groups of K pulses with each group at a different frequency. 2) At each frequency the target radar cross section is a constant during the burst but fluctuates from burst-to-burst (i.e., scan-to-scan). 3) The difference between frequencies is sufficiently large that the target radar cross section at each different frequency is independent.

Some history of this bearings-only problem at Sperry dating from 1965 is presented. The erratic estimation problems were soon recognised and labeled the 'Collapse of the Covariance Matrix Problem.' A solution was soon found as a rotation transformation on the covariance matrix. The angle of rotation was the difference between estimates of bearing before and after a measurement update. Based on a comparison of covariance analyses for simulated problems using the true bearings with Monte Carlo results it was concluded that no further significant improvement could be achieved. This conclusion is yet to be disproved.

Six experiments investigated the nature of the object-file representation supporting object continuity. Participants viewed preview displays consisting of 2 stimuli (either line drawings or words) presented within square frames, followed by a target display consisting of a single stimulus (either a word or a picture) presented within 1 of the frames. The relationship between the target and preview stimuli was manipulated. The first 2 experiments found that participants responded more quickly when the target was identical to the preview stimulus in the same frame (object-specific priming). In Experiments 3, 4, 5, and 6, the physical form of the target stimulus (a word or picture in 1 frame) was changed completely from that of either preview stimulus (pictures or words in both frames). Despite this physical change, object-specific priming was observed. It is suggested that object files encode postcategorical information, rather than precise physical information.