Comparison of Two Target Classification Techniques

It has been shown that radar returns in the resonance region carry information regarding the overall dimensions and shape of targets. Two radar target classification techniques developed to utilize such returns are discussed. Both of these techniques utilize resonance region backscatter measurements of the radar cross section (RCS) and the intrinsic target backscattered phase. A target catalog used for testing the techniques was generated from measurements of the RCS of scale models of modern aircraft and naval ships using a radar range at The Ohio State University. To test the classification technique, targets had their RCS and phase taken from the data base and corrupted by errors to simulate full-scale propagation path and processing distortion. Several classification methods were then used to determine how well the corrupted measurements fit the measurement target signatures in the catalog. The first technique uses nearest neighbor (NN) algorithms on the RCS magnitude and (range corrected) phase at a number (e.g., 2, 4, or 8) of operating frequencies. The second technique uses an inverse Fourier transformation of the complex multifrequency radar returns to the time domain followed by cross correlation. Comparisons are made of the performance of the two techniques as a function of signal-to-error noise power ratio for various processing options.     

A Sequential Approach to Target Discrimination

The Wald sequential probability ratio test is applied to the discrimination of targets observed by a radar or other sensor and a form for the classifier involving linear predictive filtering is developed. In this sequential approach, a target is illuminated with consecutive pulses until a classification of the target can be made to within a prescribed probability of error. Because of the linear-predictive formulation, the computational and storage requirements for the classifier are related only to the number of returns necessary to predict the target signature and not to the length of signature observed; a classifier with modest storage and computational requirements can be employed to process signatures consisting of an arbitrarily large number of returns. The classifier is based on some well-known results in mean-square filtering theory and has a simple intuitive interpretation. The classifier structure can also be related to autoregressive time series analysis and innovations process concepts and has an interpretation in the frequency domain in terms of the maximum entropy and maximum likelihood spectral estimates for the target signatures.     

Target classification approach based on the belief function theory

A theoretical framework is presented for target classification based on the belief theory on the continuous space. The proposed approach is applicable when class-conditioned densities of feature/attribute measurements are known only partially, as subjective models of a potential "betting" behaviour. Prior class probabilities may also be unknown. Numerical examples are provided to illustrate how the proposed approach is more cautious in decision making and produces very different results from those obtained using the Bayesian classifier.     

A technique for enhanced slant range resolution for SAR systems inknowledge-based environment

Synthetic Aperture Radar (SAR) is an airborne (or spaceborne) radar mapping technique for generating high resolution maps of surface target areas including terrain. High resolution is achieved by coherently combining the returns from a number of radar transmissions. The resolution of the images is determined by the parameters of the emissions, with more data giving greater resolution. A requirement of the Microwave Radar Division's SAR radar is to provide classification of targets. This paper presents a technique for enhancing slant range resolution in SAR images by dithering the carrier centre frequency of the transmitted signal. The procedure controls the radar waveforms so they will optimally perform the classification function, rather than provide an image of best quality. It is shown that a Knowledge-Based engineering approach to determining the waveform of the radar gives considerably improved performance as a classifier of targets (of large radar cross-section), even though the corresponding image is degraded     

Radar Detection of Correlated Targets in Clutter

This paper provides general models of radar echoes from a target. The rationale of the approach is to consider the echoes as the output of a linear dynamic system driven by white Gaussian noise (WGN). Two models can be conceived to generate N target returns: samples generated as a batch, or sequentially generated one by one. The models allow the accommodation of any correlation between pulses and nonstationary behavior of the target. The problem of deriving the optimum receiver structure is next considered. The theory of "estimator-correlator" receiver is applied to the case of a Gaussian-distributed time-correlated target embedded in clutter and thermal noise. Two equivalent detection schemes are obtained (i. e., the batch detector and the recursive detector) which are related to the above mentioned procedures of generating radar echoes. A combined analytic-numeric method has been conceived to obtain a set of original detection curves related to operational cases of interest. Finally, an adaptive implementation of the proposed processor is suggested, especially with reference to the problem of on-line estimation of the clutter covariance matrix and of the CFAR threshold. In both cases detection loss due to adaptation has been evaluated by means of a Monte Carlo simulation approach. In summary, the original contributions of the paper lie in the mathematical formulation of a powerful model for radar echoes and in the derivation of a large set of detection curves.     

Mixture Reduction Algorithms for Point and Extended Object Tracking in Clutter

The Bayesian solution to the problem of tracking a target with measurement association uncertainty gives rise to mixture distributions, which are composed of an ever increasing number of components. To produce a practical tracking filter, the growth of components must be controlled by approximating the mixture distribution. Two mixture reduction schemes (a joining algorithm and a clustering algorithm) have been derived for this purpose. If significant well spaced mixture components are present, these techniques can provide a useful improvement over the probabilistic data association filter (PDAF) approach, which reduces the mixture to a single Gaussian component at each time step. For the standard problem of tracking a point target in uniform random clutter, a Monte Carlo simulation study has been employed to identify the region of the problem parameter space where significant performance improvement is obtained over the PDAF. In the second part of this paper, the formal Bayesian filter is derived for an extended target consisting of an array of measurement sources with association uncertainty. A practical multiple hypothesis filter is implemented using mixture reduction and simulation results are presented.     

Bayesian and Dempster-Shafer target identification for radarsurveillance

This paper considers the problem of target track identification in a radar surveillance system. To build a target identifier alongside a tracker, four features which are available for real-time processing in an air surveillance system are used here: target identity (TID) from a friend-and-foe identification (IFF) system, elevation measurement from the radar, target speed, and acceleration estimated by a tracker. These four features are combined to classify air targets into five different air target categories: friendly commercial, friendly military, hostile commercial (or unknown airline), hostile military, and false targets (clutter). Two popular statistic-based techniques, namely, the Bayesian and Dempster-Shafer methods, are applied to develop radar target identification algorithms for our application. Real-life as well as simulated air surveillance radar data are used to evaluate the practicality and effectiveness of this track identification approach in a radar surveillance system     

GMM-based target classification for ground surveillance Doppler radar

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.     

Radar target identification using the bispectrum: a comparativestudy

Radar target identification is performed using time-domain bispectral features. The classification performance is compared with the performance of other classifiers that use either the impulse response or frequency domain response of the unknown target. The classification algorithms developed here are based on the spectral or the bispectral energy of the received backscatter signal. Classification results are obtained using simulated radar returns derived from measured scattering data from real radar targets. The performance of classifiers in the presence of additive Gaussian (colored or white), exponential noise, and Weibull noise are considered, along with cases where the azimuth position of the target is unknown. Finally, the effect on classification performance of responses horn extraneous point scatterers is investigated     

Radar target classification using doppler signatures of human locomotion models

The problem of target classification for ground surveillance Doppler radars is addressed. Two sources of knowledge are presented and incorporated within the classification algorithms: 1) statistical knowledge on radar target echo features, and 2) physical knowledge, represented via the locomotion models for different targets. The statistical knowledge is represented by distribution models whose parameters are estimated using a collected database. The physical knowledge is represented by target locomotion and radar measurements models. Various concepts to incorporate these sources of knowledge are presented. These concepts are tested using real data of radar echo records, which include three target classes: one person, two persons and vehicle. A combined approach, which implements both statistical and physical prior knowledge provides the best classification performance, and it achieves a classification rate of 99% in the three-class problem in high signal-to-noise conditions.     

Classification of radar targets using synthetic neural networks

Radar target classification performance of neural networks is evaluated. Time-domain and frequency-domain target features are considered. The sensitivity of the neural network algorithm to changes in network topology and training noise level is examined. The problem of classifying radar targets at unknown aspect angles is considered. The performance of the neural network algorithms is compared with that of decision-theoretic classifiers. Neural networks can be effectively used as radar target classification algorithms with an expected performance within 10 dB (worst case) of the optimum classifier     

Precision large scale air traffic surveillance using IMM/assignmentestimators

We present the development and implementation of a multisensor-multitarget tracking algorithm for large scale air traffic surveillance based on interacting multiple model (IMM) state estimation combined with a 2-dimensional assignment for data association. The algorithm can be used to track a large number of targets from measurements obtained with a large number of radars. The use of the algorithm is illustrated on measurements obtained from 5 FAA radars, which are asynchronous, heterogeneous, and geographically distributed over a large area. Both secondary radar data (beacon returns from cooperative targets) as well as primary radar data (skin returns from noncooperative targets) are used. The target IDs from the beacon returns are not used in the data association. The surveillance region includes about 800 targets that exhibit different types of motion. The performance of an IMM estimator with linear motion models is compared with that of the Kalman filter (KF). A number of performance measures that can be used on real data without knowledge of the ground truth are presented for this purpose. It is shown that the IMM estimator performs better than the KF. The advantage of fusing multisensor data is quantified. It is also shown that the computational requirements in the multisensor case are lower than in single sensor case, Finally, an IMM estimator with a nonlinear motion model (coordinated turn) is shown to further improve the performance during the maneuvering periods over the IMM with linear models     

Maximum likelihood identification of glint noise

If the non-Gaussian distribution function of radar glint noise is known, the Masreliez filter can be applied to improve target tracking performance. We investigate the glint identification problem using the maximum likelihood (ML) method. Two models for the glint distribution are used, a mixture of two Gaussian distributions and a mixture of a Gaussian and a Laplacian distribution. An efficient initial estimate method based on the QQ-plot is also proposed. Simulations show that the ML estimates converge to truths     

SAR ATR performance using a conditionally Gaussian model

A family of conditionally Gaussian signal models for synthetic aperture radar (SAR) imagery is presented, extending a related class of models developed for high resolution radar range profiles. This signal model is robust with respect to the variations of the complex-valued radar signals due to the coherent combination of returns from scatterers as those scatterers move through relative distances on the order of a wavelength of the transmitted signal (target speckle). The target type and the relative orientations of the sensor, target, and ground plane parameterize the conditionally Gaussian model. Based upon this model, algorithms to jointly estimate both the target type and pose are developed. Performance results for both target pose estimation and target recognition are presented for publicly released data from the MSTAR program     

Tracking with classification-aided multiframe data association

In most conventional tracking systems, only the target kinematic information from, for example, a radar or sonar or an electro-optical sensor, is used in measurement-to-track association. Target class information, which is typically used in postprocessing, can also be used to improve data association to give better tracking accuracy. The use of target class information in data association can improve discrimination by yielding purer tracks and preserving their continuity. In this paper, we present the simultaneous use of target classification information and target kinematic information for target tracking. The approach presented integrates target class information into the data association process using the 2-D (one track list and one measurement list) as well as multiframe (one track list and multiple measurement lists) assignments. The multiframe association likelihood is developed to include the classification results based on the "confusion matrix" that specifies the accuracy of the target classifier. The objective is to improve association results using class information when the kinematic likelihoods are similar for different targets, i.e., there is ambiguity in using kinematic information alone. Performance comparisons with and without the use of class information in data association are presented on a ground target tracking problem. Simulation results quantify the benefits of classification-aided data association for improved target tracking, especially in the presence of association uncertainty in the kinematic measurements. Also, the benefit of 5-D (or multiframe) association versus 2-D association is investigated for different quality classifiers. The main contribution of this paper is the development of the methodology to incorporate exactly the classification information into multidimensional (multiframe) association.     

Multiple model particle flter track-before-detect for range ambiguous radar

The middle pulse repetition frequency（MPRF）and high pulse repetition frequency（HPRF）modes are widely adopted in airborne pulse Doppler（PD）radar systems,which results in the problem that the range measurement of targets is ambiguous.The existing data processing based range ambiguity resolving methods work well on the condition that the signal-to-noise ratio（SNR）is high enough.In this paper,a multiple model particle flter（MMPF）based track-beforedetect（TBD）method is proposed to address the problem of target detection and tracking with range ambiguous radar in low-SNR environment.By introducing a discrete variable that denotes whether a target is present or not and the discrete pulse interval number（PIN）as components of the target state vector,and modeling the incremental variable of the PIN as a three-state Markov chain,the proposed algorithm converts the problem of range ambiguity resolving into a hybrid state fltering problem.At last,the hybrid fltering problem is implemented by a MMPF-based TBD method in the Bayesian framework.Simulation results demonstrate that the proposed Bayesian approach can estimate target state as well as the PIN simultaneously,and succeeds in detecting and tracking weak targets with the range ambiguous radar.Simulation results also show that the performance of the proposed method is superior to that of the multiple hypothesis（MH）method in low-SNR environment.     

JEM modeling and measurement for radar target identification

The jet engine modulation (JEM) phenomenon, observed in radar returns from the rotating structure of jet engines, has been successfully exploited for aircraft target identification in a number of experimental radar systems. The authors develop a parametric model based on the periodic modulation of the scattered return, motivated by the potential reduction in time-on-target for reliable target identification provided by parametric models as well as by gaining insight into the JEM phenomenon. They compare the model with JEM measurements made with an experimental radar system and discuss the implications for JEM-based target identification systems     

Adaptive Detection Algorithms for Multiple-Target Situations

The performance of a mean-level detector is considered for the case where one or more interfering target returns are present in the set of cells used in estimating the clutter-plus-noise level. A serious degradation of detection probability is demonstrated for all of the single-pulse Swerling target fluctuation models (i. e., cases 0, 2, and 4). Indeed, for fixed mean radar cross sections of the primary and interfering targets, the probability of detecting the primary target is asymptotic to values significantly less than unity as the signal-to-noise ratios of the returns approach infinity. A class of alternative adaptive detection procedures is proposed and analyzed. These procedures, based on ranking and censoring techniques, maintain acceptable performance in the presence of interfering targets, and require only a minor addition in hardware to a conventional mean-level detector.     

Adaptive learning approach to landmine detection

We consider landmine detection using forward-looking ground penetrating radar (FLGPR). The two main challenging tasks include extracting intricate structures of target signals and adapting a classifier to the surrounding environment through learning. Through the time-frequency (TF) analysis, we find that the most discriminant information is TF localized. This observation motivates us to use the over-complete wavelet packet transform (WPT) to sparsely represent signals with the discriminant information encoded into several bases. Then the sequential floating forward selection method is used to extract these components and thereby a neural network (NNW) classifier is designed. To further improve the classification performance and deal with the problem of detecting mines in an unconstraint environment, the AdaBoost algorithm is used. We integrate the feature selection process into the original AdaBoost algorithm. In each iteration, AdaBoost identifies the hard-to-learn examples and a new set of features which provide the specific discriminant information for these hard samples is extracted adaptively and a new classifier is trained. Experimental results based on measured data are presented, showing that a significant improvement on the classification performance can be achieved.