This paper addresses small targets and signal processing from the perspective of rejecting radiation noise spikes. Nuclear and space radiation create noise spikes inside infrared detectors causing an overwhelming number of false alarms, if steps are not taken to mitigate the radiation noise spikes. Traditional radiation device/circuit hardening methods are effective, but must be reapplied to each new technology forcing special design point solutions and parts that are increasingly economically nonviable. Real-time noise mitigation algorithms represent a general hardening solution and have been demonstrated for both interceptor seeker and space surveillance sensor applications. A new combined HWIL/radiation synthetic test environment has been developed that enables real-time algorithm evaluation over the total system performance envelope, under flight motion simulation and fully dynamic optical sensor scene stimulation. This work was sponsored by the Defense Special Weapons Agency.

This paper proposes a new algorithm which is able to decompose the measurements gathered within the same frame into multiple resolutions. The decomposition is based on the geometric relation between the measurements. The statistics of the decomposed measurements are derived with which multiresolutional models are constructed. The computation of the statistics is approximated so that it is efficient in computation. Simulations are designed to analyze the performance of the algorithm and to illustrate the significance of this multiresolutional approach. The results of the simulations prove that multiresolutional data and model structures can significantly reduce the computational burden of the tracking algorithms while the performance of a multiresolutional tracker is still comparable to that of a conventional one.

We present two algorithms for determining sensor motion of a multi- spectral push-broom imager for use in subsequent image registration. The first algorithm, termed the 'pairwise' method, performs cross-correlations between individual pairs of channels. The offsets of maximum correlation are formulated into a system of linear equations whose solution gives an estimate of the jitter function. The second algorithm performs cross-correlations between channels and a reference image called the 'baseline' which is constructed by averaging together all the channels in the image cube. An estimated jitter time series is computed for each channel, all of which are overlapped and averaged to obtain a best estimate of the jitter function. The pairwise method is more general in that it can handle a wider range of jitter scenarios. The baseline method, although more restricted, is very simple to implement, and its accuracy can be improved substantially through iteration. In this paper, we describe both methods in detail and present results of simulations performed on thermal-infrared data cubes.

Sea clutter amplitude is often modeled as a compound random variable Z equals AX, where A is a positive valued random variable and X has a Rayleigh distribution. The K, class A, and discrete Rayleigh mixture distributions can be derived from these assumptions. Moreover, successive values of A may be correlated. If A is modeled as a finite Markov process, Z is described by a hidden Markov model (HMM). The applicability of Rayleigh mixture and hidden Markov models to RADAR sea clutter is demonstrated empirically. Amplitude only and phase coherent detection statistics are derived from these models using locally optimal and likelihood ratio techniques. Robust implementations of the locally optimal processor based on the Rayleigh mixture model have been developed, and empirical ROC curves demonstrate performance improvement of up to 9 dB in comparison with a CFAR detector for small targets in sea clutter. In a test case, the locally optimal hidden Markov detector is then shown to offer an additional 3 dB over the Gaussian mixture detector. Further examples compare the amplitude and phase coherent hidden Markov detectors with CFAR and Doppler processors.

A method of detecting dim moving targets in highly cluttered time- varying image sequences is presented. These targets are sufficiently faint that they cannot be detected in individual image frames, so a track-before-detect method is required. This method is based on the analysis of a single 2-D composite image in which movement shows up as bright swaths. The composite image is constructed by computing the absolute values of the differences between neighboring image frames and then, for each pixel, taking the maximum value of these differences over time. Image registration is employed where required. Moving targets are detected using a wavelet-based feature extraction algorithm, in which the bright regions in the composite image are decomposed into bar features. Background rejection is employed to remove sensor noise and other changes in brightness not associated with movement, for example changes in illumination. Initial clutter rejection is based on a given threshold exceedence rate estimated from the multiresolution background statistics of feature intensity. Further false track rejection involves a space-time analysis of image frame differences in the neighborhood of each bright region, where genuine movement appears as connected diagonal swaths. Results so far indicate that this approach to moving target detection is highly effective when backgrounds are highly correlated in time. No assumptions about the behavior of the target motion are required, the only limitation being that the targets are moving sufficiently quickly to traverse several pixels during the total time interval under analysis.

In many remote sensing techniques the measured signal can be modelled as the result of a convolution operator (with completely or partially known impulse response) on an input signal which is known to be the superposition of a finite number of elementary signals with unknown parameters. The restoration or inversion problem becomes then the estimation of these parameters. In this work we propose a Bayesian estimation framework to solve these inverse problems by introducing some prior knowledge on the unknown parameters via the specified prior probability laws on them. More specifically, we propose to use the maximum a posteriori (MAP) estimation method with some specific choices for the prior laws. The MAP criterion is optimized using a modified Newton-Raphson algorithm. Some simulation results illustrate the performances of the proposed method. In these simulations we considered the input signal to be the superposition of Gaussians with unknown positions, standard deviations and amplitudes.

Questions of detector synthesis based on the generalized signal processing algorithm for signals with random initial phase are considered. Variance and variance estimation of total noise component at the generalized detector output under the finite time interval [0, T] are determined. Comparative analysis of detection characteristics of the optimal and generali detectors is carried out. Investigation avenues to stabilization of detection characteristics under employment of the generalized detector independent of signal random parameters based on phase tracking system have been proposed.

A track-before-detect processor has been developed by Litton Data Systems Division (DSD), Agoura Hills, Calif. and implemented in hardware. Its objective was to enhance radar and IR dim target detection in adverse conditions. The unit has been integrated with an operational radar system and tested in real time scenarios. This paper describes the processor target detection performance in sea clutter, weather and noise. In specific it addresses the issue of maritime radar detection performance as were observed during sea trials data collection tests. Performance in sea clutter for infra red (IR) sensor is also provided.

The velocity filter follows naturally from the derivation of the 3D (spatio-temporal) matched filter. It assumes constant velocity targets and that the spatial distribution of the targets remains unchanged. Typically the target velocity is unknown and a bank of velocity filters is implemented which covers the range of possible target velocities. Classically it is assumed that the gain of the velocity filter is proportional to the square root of the number of image frames filtered. As the number of frames is increased to enhance the filter gain, the width of the filter response in velocity space decreases. Consequently, more gain requires more filters to cover the same range of velocities. Each of the filters produces a set of false alarms. To maintain a constant false alarm rate, the detection threshold must be increased. The higher detection threshold reduces the probability of detection and offsets some of the gain achieved by the velocity filters. This performance trade is quantitatively analyzed.

The ability to detect weak targets of low contrast or signal-to- noise ratio (SNR) is improved by a fusion of data in space and wavelength from multispectral/hyperspectral sensors. It has been demonstrated previously that the correlation of the clutter between multiband thermal infrared images plays an important role in allowing the data collected in one spectral band to be used to cancel the background clutter in another spectral band, resulting in increased SNR. However, the correlation between bands is reduced when the spectrum observed in each pixel is derived from a mixture of several different materials, each with its own spectral characteristics. In order to handle the identification of objects in this complex (mixed) clutter, a class of algorithms have been developed that model the pixels as a linear combination of pure substances and then unmix the spectra to identify the pixel constituents. In this paper a linear unmixing algorithm is incorporated with a statistical hypothesis test for detecting a known target spectral feature that obeys a linear mixing model in a mixture of background noise. The generalized linear feature detector utilizes a maximum likelihood ratio approach to detect and estimate the presence and concentration of one or more specific objects. A performance evaluation of the linear unmixing and maximum likelihood detector is shown by comparing the results to the spectral anomaly detection algorithm previously developed by Reed and Yu.

Precision registration (alignment) of images is utilized in dim target detection, temporal change detection, and other surveillance applications. The Cramer-Rao bound on the accuracy of parameter estimation governs the fundamental limit on the accuracy of image registration. This presentation derives the Cramer-Rao bound on registration accuracy, with emphasis on how registration accuracy depends on sensor noise, scene geometry, image characteristics and the number of pixels used in the registration process, and compares the Cramer-Rao bound prediction to registration results from the SAIC IR processing code, STAS, for both synthetically generated scenes and real IR data.

In this paper we present a new technique for the estimation of the velocity of moving targets using sequential frames. This estimation process may be used to estimate a potential set of velocities of moving targets which in turn may be used by three-dimensional (3-D) directional matched filters. It may also be used as a target trajectory estimation technique. The method is based on a local probability density matching segmentation technique with spatiotemporal associations. Experimental results are presented.

A pipe-lined algorithm and parallel architecture is under development for real time detection of sparse small objects in images. Monochromatic images from an airborne active infrared scanner, images from a low-altitude aircraft-mounted multispectral scanner, and passive infrared imagery obtained from cameras mounted on ground vehicle are the image types intended for the application of this system to the detection of minefields. The paper briefly describes the characteristics of these three different kinds of image sensors and the operating environments. The general image processing system architecture and the functions of each of the components are also presented. The feature selection and algorithm adaptations for each of the image classes are described. Preliminary results obtained from an experimental system consisting of a small transputer network and array processors are discussed.

The synthetic theater of war (STOW) is the major application of a Defense Advanced Research Projects Agency (DARPA) thrust in advanced distributed simulation (ADS). The STOW Program focuses on an advanced concept technology demonstration (ACTD) termed STOW 97 sponsored by DARPA with the United States Atlantic Command (USACOM). The successful implementation of STOW 97 technologies in November 1997 with the United Endeavor 98-1 exercise will mark the full operational capacity of the USACOM Joint Training, Analysis and Simulation Center. To support ADS applications up to the Joint Task Force level, STOW seeks to develop and demonstrate technologies enabling the integration of war-fighting through virtual and constructive simulations from geographically distributed locations in a common synthetic battlespace.

The need for a high-fidelity sensor design simulation model to accurately predict the system performance envelope and to offset the escalating cost of the system development and testing is widely accepted by the defense community. This paper presents one such example of the modeling capability developed for the ballistic missile defense (BMD) application, called the signal processing environment for analysis and reduction (SPEAR) simulation. SPEAR has become a key IR sensor design and signal processing performance verification tool for the BMD Advanced Sensor Technology Program (ASTP), the Discriminating Interceptor Technology Program (DITP), and the ground based interceptor (GBI) and, where it is used for sensitivity analyses, algorithm evaluations, and performance assessments. For these programs, SPEAR provides an algorithm testing simulation to evaluate candidate signal processing options, and implement and test performance of algorithms proposed through advanced technology programs. In addition, SPEAR is used to process real world data to provide assessments of sensor performance and provide preflight predictions. The simulation has been interfaced to the synthetic scene generation model (SSGM), a community standard background and target scene generation simulation. Through this interface sensor performance can be evaluated against realistically modeled backgrounds to evaluate filtering, detection, and false alarm performance. SPEAR is a hi-fidelity passive infrared (IR) sensor and signal processing simulation for staring scanning, and hybrid sensors. It allows the user to specify the IR sensor physics including the sensor, optics, focal plane array or scan chip assembly, analog signal processor, time dependent and object dependent processing parameters and specific noise sources such as optics, jitter, fixed pattern noise, dark current, and gamma spike noise. SPEAR is an Ada/PVWAVE combination. The sensor and signal processing is written in Ada and the execution, parameter input, and function analysis are controlled with the graphical user interface (GUI) written in PVWAVE. The signal processing techniques available as options include time dependent processing techniques such as adaptive threshold detection, background estimation and removal, morphological filtering, match filtering, target signature extraction, and object dependent processing techniques such as centroiding and pulse matching. SPEAR has simulation control options to allow the user to execute and examine data per frame (mission mode) or in a statistical mode to investigate parametric sensitivities of the sensor performance. Documentation of SPEAR includes manuals on the GUI, the SPEAR application components, and guidelines for adding new algorithms and features. This paper provides a summary of key algorithm and options in SPEAR. Examples of performance analysis results are provided. The paper includes stochastic analyses of both the above- the-horizon and below-the-horizon engagements of target and background generated scenes using SSGM. Also discussed are the evaluation of radiometric measurement precision, angular measurement precision, and detection of targets of varying intensities with respect to varying sensor signal processing techniques.

In this paper, a method of testing combinations of image processing, track-before-detect, and track-after-detect algorithms is presented. It emphasizes false track confirmation rates and the time required to confirm true tracks whereas methods in the literature emphasize track purity and accuracy in estimating the target state. This method, which is an extension of the Neyman- Pearson criterion, yields a single performance measure, the expected time to confirm a true target track. The value of this method is in potential for component algorithm (spatial filter, track-before-detect, track-after-defect) tradeoff and track discriminant studies. Using it, one may quantitatively compare the effect of different spatial filters or different track discriminants, or the effect of using more computationally intensive track-after-detect algorithms and less computationally intensive track-before-detect algorithms.

Parameter estimation can be performed by either optical/digital correlation techniques or digital adaptive filtering. For a three- dimensional estimation of parameters, the problem is difficult to solve with correlation techniques, essentially because the correlation of the observed image with all reference images of the database is not practically possible. We propose a solution based on Kalman filter-aided correlation. This reduces the number of required correlations. The parameter estimation resulting from the digital filtering is used as the a priori knowledge. The correlation of the observed image is then performed with a small number of correlation filters, selected from the a priori information. The a posteriori parameters are given by whichever of the selected correlation filters provides the maximum correlation function. The method is tested in the particular application of determining the 3-D attitude of maneuvering targets.

Star sensors developed in the last years can be enhanced in terms of mass reduction, lower power consumption, and operational flexibility, by taking advantage of improvements in the detector technology and the electronics components. Jena-Optronik GmbH developed an intelligent modular star and target named 'stellar and extended target intelligent sensor' (SETIS). Emphasis was placed to increase the sensor adaptability to meet specific mission requirements. The intelligent modular star and target tracker shall generate positional information regarding a number of celestial targets or shall act as a navigation camera. The targets will be either stars or extended objects like comets and planetary objects, or both simultaneously. Deign drivers like simultaneous tracking of extended targets and stars or searching for new objects during tracking of already detected objects require a powerful hard-wired digital data preprocessing. An advanced rad-tolerant ASIC- technology is used for the star tracker preprocessor electronics. All of the necessary preprocessing star tracker functions like pixel defect correction, filtering, on-line background estimation, thresholding, object detection and extraction and pixel centroiding are realized in the ASIC design. The technical approach for the intelligent modular star and target tracker is presented in detail. Emphasis is placed on the description of the powerful signal preprocessing capabilities.

In a multitarget environment, tracking systems must include methods for associating measurements to targets. The complexity of that task is compounded when data from multiple sensors is available. This paper presents a clustering approach to the multitarget multisensor tracking problem. The measurement set is partitioned into equivalence classes (clusters) and the data association problem is redefined to be one of associating the cluster centers and the tracks, resulting in a significant reduction in the size of the association problem. Track termination and track initiation are part of system design, therefore allowing the designed system to be tested on elaborate multitarget tracking scenarios involving an unknown and changing number of real aircraft trajectories. Methods for evaluating the performance of the tracking system, as well as the clustering algorithms are introduced. An equivalence relation clustering algorithm is derived and compared by Monte-Carlo simulations to the subtractive clustering algorithm. The tracking system is shown to effectively track seven crossing aircraft trajectories of different duration, in the presence of clutter. Track maintenance is performed by extended Kalman filters.

A standard assumption of most multiple target tracking filters is that all the targets move independently of one another. However, in many cases, it is known a-priori that the targets move (at least approximately) as a group: this dependency should not be ignored. In this paper we describe an approach to multiple target tracking where the target dynamics are taken to be the superposition of a group effect which is common to all group members and an individual effect which is taken to be independent between members of the group. The method also allows for the presence of clutter and missed target detections. This is done by embedding the dependent target motion model within a multiple hypothesis framework. A closed form solution is derived for the special linear-Gaussian case and simulation results illustrating performance are presented. This paper is a recursive extension of the selection method presented at last year's conference.

The concept of retrodiction of discrete probabilities is exploited in this paper to provide alternative data association algorithms for tracking multiple targets with a single or multiple sensors. These algorithms employ multiple frames of data in the data association processing. The concept of retrodiction is also applied to the task of multiple model filtering. Alterative optimization criteria are also exploited to provide alternative association approaches and each approach is expected to exhibit different estimation error characteristics. These additional association approaches provide wider selection to better tailor the tracking algorithms to a specific application. These approaches offer improved performance over single-frame association tracking approaches. This improved performance is obtained, however, at the expense of increased processing load. A number of different approaches are described that employ multiple-frame data association. Similarly, a number of different approaches are also described that employ a moving window of multiple measurements for multiple model filtering. With these algorithms, design parameters can be selected to adjust performance to suit a specific application.

In distributed track-level level fusion system, it is a well-known fact that successful fusion of tracks from offboard sources requires that these tracks do not contain underlying biases or offsets. Unfortunately, this lack of bias or offset is often not the case, because the offboard tracking system references its tracks to a coordinate system that is offset and misaligned with respect to truth due to navigational drift and sensor misalignment. In this paper, we present a technical approach for both detecting and correcting for these biases in a noncooperative target sense. Furthermore, the algorithms are configured to operate in a multiple-hypothesis tracking environment. These algorithms have been implemented in a simulated air threat environment, and performance improvements have been noted of up to an order of magnitude in target/track miss distance.

A general theory of multi-object state-estimation problems, also known as multi-target tracking problems, is presented, using explicit random-set formalism. Probability density functions of random sets, as well as Choquet's capacity functionals, are used to represent random sets, in pursuit of the possibility of such a theory becoming a theoretical foundation of data fusion theory. The theoretical and algorithmic developments over the past three decades in this area are also re-examined in the light of this new formalism, as well as the recent development of correlation-free algorithms that utilize random-set formalism explicitly.

The reduced sufficient statistics (RSS) algorithm was originally developed by Kulhavy for parameter estimation only. Here, we present a modified form of the RSS algorithm which recursively computes a model posterior probability density function for the state vector. The model density is chosen to be a multidimensional Haar basis representation with dyadic scale, such that the basis functions are disjoint hypercubes. It is then shown that the model density coefficients can be obtained in closed form for this choice of basis functions. A critical part of the modified RSS algorithm is the approximation of the one-step predicted density for the state vector. It is shown that when the transition density for the state vector also has dyadic scale, that a closed-form recursion is ultimately obtained for both predicted and filtered approximating densities. Finally, an application of the algorithm to target tracking using bearings-only measurements is given.

This paper presents a detailed discussion of clustering as applied to multiple hypothesis tracking (MHT). The combinatorial problem associated with forming multiple data association hypotheses can be reduced significantly by partitioning the entire set of system tracks and input data elements into separate clusters. Cluster management, a process that deals with cluster formation, merging, splitting and deletion, is thus motivated by the idea that a large tracking problem can be divided into a number of smaller problems that can be solved independently. The paper emphasizes on the cluster splitting process since it is the most difficult aspect of clustering while being an often neglected issue in the target tracking literature. The hypothesis dependencies that must be taken into account when one attempts to split the hypothesis tree of a cluster into two or more independent trees are discussed. This is an important issue since the hypotheses within a cluster must not interact with the hypotheses contained within other clusters for the MHT technique to remain consistent. A very efficient algorithm is described that performs a combined split-merge process simultaneously for all the clusters. The algorithm has been designed to avoid a waste of computer resources that may happen when splitting clusters that should have been kept merged according to the most recent input data set. The dynamic data structure that is used to implement the hypothesis tree is described as a key element of the approach efficiency. An example of cluster management is presented.

This paper addresses the problem of the estimation of the trajectory of a tactical ballistic missile using line of sight (LOS) measurements from one or more passive sensors (typically satellites). The major difficulties of this problem include: the estimation of the unknown time of launch, incorporation of (inaccurate) target thrust profiles to model the target dynamics during the boost phase and an overall ill-conditioning of the estimation problem due to poor observability of the target motion via the LOS measurements. We present a robust estimation procedure based on the Levenberg-Marquardt algorithm that provides both the target state estimate and error covariance taking into consideration the complications mentioned above. An important consideration in the defense against tactical ballistic missiles is the determination of the target position and error covariance at the acquisition range of a surveillance radar in the vicinity of the impact point. We present a systematic procedure to propagate the target state and covariance to a nominal time, when it is within the detection range of a surveillance radar to obtain a cueing volume. Mont Carlo simulation studies on typical single and two sensor scenarios indicate that the proposed algorithms are accurate in terms of the estimates and the estimator calculated covariances are consistent with the errors.

The integration of multiple sensors for target tracking and resource management has been intensely investigated and several effective techniques have been developed. These conventional techniques employ decision-directed logic and are very complex but have the potential to improve performance. For most systems, each sensor provides its information to a central location where the integration occurs. The central track is employed for system decisions and it is typically not used by the individual sensors. This low level of integration provides a manageable tracking environment but restricts the potential for system improvement. An electronically scanned array (ESA) is highly controllable and has the ability to greatly enhance tracking performance. Resource allocation for an ESA is critical since it must support multiple functions, and several modern techniques have been developed to enhance its performance as a stand-alone sensor by effectively managing its time-energy budget. The integration of an ESA with other sensors can further enhance the tracking and reduce the resource allocation requirements of the ESA. This paper presents a technique for ESA resource management through the use of multisensor integration. The proposed technique avoids the decision-directed logic associated with conventional techniques by employing the interacting multiple model (IMM) algorithm. Simulation results are provided to demonstrate the effectiveness of this modern integration technique.

Large classes of data association problems in multiple hypothesis tracking applications involving multiple and single sensor systems can be formulated as multidimensional assignment problems, which are then fundamental in both track initiation and maintenance. In particular, track maintenance uses an N-frame moving window over the scans of data from multiple sensors giving rise to the repeated solution of (N plus 1)-dimensional assignment problems. Due to the complexity of these NP-hard problems and the fact that one must solve these problems to the noise level in the problem, it is most advantageous to take advantage of a previously solved problem over a previous window. This work explores this issue in track maintenance.

This paper extends the two set data association performance model developed by Mori, et al to include miss detections and bias. The referenced paper developed an analytical model for the probability of correct association of two data sets, called 'tracks' and 'measurements,' using an optimal 2 dimensional assignment algorithm, where the 'true' objects are distributed uniformly but at random in a circular disk. For these true objects, measurements are obtained by adding independent random errors with the same covariance. Tracks are obtained in the same way except a different, fixed covariance is used. Finally, one of the data sets includes an additional distribution of random points, considered 'false alarms.' This paper extends their results to obtain an analytical model that accounts for bias between the data sets and missed detections in either data set. The analytical model is useful in assessing the impact of system requirements for sensor sensitivity, random error and inter-sensor bias error on measurement-to- measurement, measurement-to-track or track-to-track association.

Over-the-horizon skywave radar exploits ionospheric propagation of HF signals to detect targets beyond the line-of-sight horizon. Multiple propagation paths between the radar sites and the target are often encountered, giving multiple resolved detections for a single target. An algorithm for the fusion of multipath tracks is outlined here which accounts for uncertainty in the coordinate registration transformation to ground coordinates. A multihypothesis track association procedure is described which may be appended to existing radar coordinate tracking filters. The probability for each feasible track association hypothesis is computed, and fused estimates for target states in ground coordinates are evaluated for each hypothesis.

In this paper we present the development and implementation of a multisensor-multitarget tracking algorithm for large scale air traffic surveillance based on the IMM state estimator combined with a 2-dimensional assignment for data association. The algorithm can be used to track a large umber 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 non-cooperative 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 the IMM estimator is compared with that of the Kalman filter. 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 Kalman filter. 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.

A variable-structure multiple-model (VSMM) estimator, called model- group switching (MGS) algorithm, has been developed recently. It is the first VSMM estimator that is generally applicable to a large class of problem with hybrid (continuous and discrete) uncertainties. In this algorithm, the model set is made adaptive by switching among a number of predetermined groups of models. It has the potential to be substantially superior to fixed-structure MM estimators, including the interacting multiple-model (IMM) estimator. Many issues in the application of this algorithm are investigated, such as the model-group activation logic and model- group design, via a detailed design for a problem of tracking a maneuvering target using a time-varying set of models, each characterized by a representative value of the target's expected acceleration. Simulation results are given to demonstrate the performance (based on reasonable and complete measures) and computational complexity of the MGS algorithm, relative to the IMM estimators, under carefully designed random and deterministic scenarios.

The method of virtual measurements (MOVM) is described for designing nonlinear filters. The new nonlinear filter theory generalizes the Kalman filter, and in some important applications, the performance of the new filter is vastly superior to the extended Kalman filter (EKF). Unlike the EKF, the new theory does not use linearization. The new design approach, MOVM, can be applied to exact nonlinear filters as well as nonlinear approximate filters. Several examples are given in which the performance of the new nonlinear filter is compared with the extended Kalman filter.

In contrast to situations with well-separated or crossing targets, two phenomena make tracking of closely-spaced objects more difficult: First, the data association problem for multiple targets must be handled over a longer time until the group finally dissolves into well-separated targets. The critical moment of split-off is typically unknown and may even be accompanied with strong maneuvers. Second, due to the finite resolution of every physical sensor, closely-spaced targets might well be unresolvable. With a simplified model of the sensor resolution, possibly unresolved measurements can be treated in combination with the association task. The problems addressed are even more significant in case of dense clutter interference, imperfect target detections, and inaccurate measurements. By taking resolution conflicts explicitly into account, we discuss quantitative results for track maintenance under various conditions within the framework of Bayesian IMM-MHT methods, i.e. tracking of maneuvering formations during spatially or temporarily limited interference. The filter performance is evaluated for typical radar parameters and air situations involving two targets. In particular, the phenomenon of mixing target identities during a formation flight is addressed. The achieved results are compared with more conventional alternatives (MHT, standard JPDAF) that do not consider resolution conflicts.

Interacting multiple model (IMM) filtering and multiple hypothesis tracking (MHT) represent the most accurate methods currently available for tracking multiple maneuvering targets in cluttered environments. Although these methods are complex, modern computational capabilities make their combined implementation feasible for modern tracking systems. This paper discusses alternative approaches for developing a combined IMM/MHT tracking system and describes specific implementations that have been developed for radar and IRST applications. Simulation results for both radar and IRST systems are presented. Track maintenance and accuracy performance for the IMM/MHT system is compared with that obtained from an MHT tracker using a conventional filtering approach. Results indicate that the improvements in data association derived from the use of IMM filtering with MHT may be comparable to the well known IMM improvements in track accuracy. The addition of multiple filter models to MHT data association has the potential to significantly increase computational requirements. Thus, several compromises, described in this paper, have been developed in order to assure computational feasibility. Preliminary estimates of computational requirements are given in order to demonstrate implementation feasibility.

In this paper we present a new technique for data association using multiassignment for tracking a large number of closely spaced (and overlapping) objects. The algorithm is illustrated on a biomedical problem, namely the tracking of a group of fibroblast (tissue) cells from an image sequence, which motivated this work. The algorithm presents a novel iterated approach to multiassignment using successive one-to-one assignments of decreasing size with modified costs. The cost functions, which are adjusted depending on the 'depth' of the current assignment level and on the tracking results, are derived. The resulting assignments are used to form, maintain and terminate tracks with a modified version of the probabilistic data association filter, which can handle the contention for a single measurement among multiple tracks in addition to the association of multiple measurements to a single track. Estimation results are given and compared with those of the standard 2-dimensional one-to-one assignment algorithm. It is shown that iterated multiassignment results in superior measurement-to- track association.

When two targets are closely-spaced with respect to the resolution of a radar, the measurements of the two targets will be merged when the target echoes are not resolved in angle, range, or radial velocity (i.e., Doppler processing). Monopulse processing is considered for direction-of-arrival (DOA) estimation of two unresolved Rayleigh targets with known relative radar cross section (RCS). The probability distribution of the complex monopulse ratio is developed for two unresolved Rayleigh targets. The Fisher information matrix and Cramer Rao bounds are developed for the DOA estimation of two unresolved Rayleigh targets using a standard monopulse radar. When the two Rayleigh targets are separated by more than one-half of the radar beamwidth, DOA estimation is accomplished for each target by treating the other target as interference. When the two targets are separated by less than one- half of a beamwidth, the antenna boresight is pointed between the two targets, and the mean of the in-phase (i.e., the real part) monopulse ratio and the variance of the in-phase and quadrature monopulse ratios are utilized to estimate the DOAs of the two targets. Simulation results that illustrate performance of the DOA estimators are given along with a simple tracking example.

Tracking for airborne early warning (AEW) weapon systems present a number of formidable challenges for any tracking and data fusion algorithms. Realistic scenarios involve thousands of targets in highly cluttered environments with multiple sensors. The E-2C weapon system must detect, track and identify these targets in as small a time frame as possible. As part of ongoing E-2C advanced tracking algorithm development activities a novel approach has been developed that utilizes the debiased coordinate conversion filter developed by Bar-Shalom and Lerro (1993) for range, and azimuth angle processing from the radar and standard EKF for rdot and other angular measurements from other sensors identified as a combined Kalman filter (CKF). To solve the data association problem the JVC algorithm [Jonker-Volgenant Castanon (1988)] was chosen because of favorable results from published studies and internally conducted in-house studies that demonstrate its speed and efficiency in solving the assignment problem for sparse matrices which is typical for E-2C applications. Results shown are based on a scenario consisting of 120 straight line and maneuvering targets overlaid on a previously recorded dense radar environment. Future plans have been initiated to incorporate other sensors and consider other association algorithms such as multi-hypothesis tracking (MHT) or interactive multiple model joint probabilistic data association filter (IMMJPDAF).

Fully adaptive matched filters typically can suppress clutter to the level of the sensor fixed pattern noise. A fully adaptive filter assumes that the clutter is a wide-sense stationary process which can be modeled by a constant means and unknown covariance function. Fixed pattern noise within a data sequence is unknown and tends to be a non-stationary process. As a result fixed pattern noise is minimally affected by fully adaptive filters. The signal processing philosophy for detecting unresolved targets is to enhance the target signal based on the sensor point spread function. When sensor fixed pattern noise exists, the signal from a point target can be significantly different from the sensor point spread function and can result in a loss in SCR. This SCR loss can make weak targets undetectable. This paper describes the effect of a fully adaptive filter on fixed pattern noise manifested as channel dependent bias and gain errors. Spectral analysis which quantifies the impact of these errors is presented. Experimental results on synthetic data and on real data from an infrared scanning sensor with channel dependent fixed pattern noise are given.

When tracking a target in clutter, a measurement may have originated from either the target, clutter, or some other source. The measurement with the strongest intensity (amplitude) in the neighborhood of the predicted target measurement is known as the 'strongest neighbor' (SN) measurement. A simple and commonly used method for tracking in clutter is the so-called strongest neighbor filter (SNF), which uses the SN measurement at each time as if it were the true one. This paper presents analytic results, along with discussions, for the SN measurement, including the a priori and a posteriori probabilities of data association events and the conditional probability density functions. These results provide theoretical foundation for performance prediction and development of improved tracking filters.

For track maintenance, there are primarily three generic sensor data fusion algorithm architectures, namely, central fusion, track fusion, and what will be referred to as composite measurement fusion. In central fusion, the sensor measurements are distributed by each sensor and the measurements from multiple sensors are then used to update the global tracks. Measurements are also called: returns, observations, threshold exceedances, plots, contacts, or hits depending on the sensors involved. In contrast, in track fusion a sequence of measurements is processed at the sensor or platform level to form tracks that are then distributed and this track data is then used to update the global tracks. Track fusion is also sometimes called hierarchical fusion, federated fusion or distributed fusion. Finally, in composite measurement fusion the measurements from multiple sensors for each apparent target are first combined to form a composite measurement and then the composite measurements are then used to update the global tracks. The tracks typically include features or other target classification information. Each of these algorithm architectures has their own advantages and disadvantages. For example, track fusion may lead to substantially reduced communications load and that can be important for physically distributed platforms. Track fusion also tends to be less sensitive to residual sensor bias errors. Central fusion typically provides more timely information. Also, for certain types of target scenarios and sensor suites, central fusion provides better accuracy in both estimation and target classification. Recent developments in track fusion make a particular hybrid fusion algorithm architecture not only appealing but practical. In this hybrid fusion, either measurement or track data in the form of a tracklet is distributed from a sensor for a target. This approach offers many of the advantages of both the centralized and track fusion algorithm architectures. This paper describes a specific hybrid fusion algorithm architecture, the decision logic for distributing measurement or track data, and the recent track fusion innovations that make this hybrid fusion practical.

Many multi-target tracking systems work by continually updating the target tracks on the basis of received target measurements. This involves solving the assignment problem, i.e. finding the maximum set of track-to-measurement associations with the largest overall likelihood. To minimize the number of association errors, it is traditional to include targets for which there are no corresponding tracks and tracks for which there are no corresponding measurements. These missing tracks and missing measurements are taken into account by augmenting the track-to-measurement likelihood matrix. In this paper, we present a new approach to solving the assignment problem which does not involve augmenting the likelihood matrix. The solution found in the new approach contains n* high-quality track-to-measurement associations. This set of n* associations optimally satisfies two criteria: (1) a high average likelihood and (2) n* close to the expected number of true track-to-measurement associations. Thus, the number of track-to- measurement associations, n*, is not specified beforehand but rather is an output of the algorithm. The two criteria are defined by fuzzy membership functions, and the solution is found using a fuzzy multi-criteria decision-making algorithm.

Many solutions have been proposed to solve the multi-target tracking problem, using location informations provided by a radar sensor. The major drawbacks of classical methods are: (1) they need a priori knowledge on the problem, such as the measurement noise statistics, the number of targets, the probabilities of detection and false alarm; (2) they only treat one part of the problem, that is to say the plot/track association problem; (3) they are unable to solve the initiation problem, which is the key problem. We recently developed a neural solution for multiple radar target tracking, allowing to solve the problem in an integrated way, with few assumptions on the input data. This paper presents how measurement attributes, such as the Doppler velocity, the detection likelihood (classification probability of the detection) and the local report density (estimated on-line can be incorporated in our neural solution. Some simulation results are reported to compare the performances of our neural solution (with and without measurement attributes), with the joint probabilistic data association filter.

This paper presents and compares two alternative classes of solutions to the plot-to-track association problem. The first class of solutions relies on classical approaches of signal processing, principally based on the Bayes theory, that is to say the nearest- neighbor filter, the probabilistic data association filter and the joint probabilistic data association filter. The data association problem can be reduced to a combinatorial optimization problem, for which the time needed to obtain the exact solution grows drastically with the problem size. This is the reason why, in most cases, we do not look for the best solution, but rather for a good solution, reachable in a reasonable computation time. Consequently, neural networks are an interesting alternative to classical solutions. We first review several neural models: Hopfield networks, Boltzmann machine, mean filed approximation networks and our approach derived from the Hopfield model. Then we present some simulation results that enable to compare the various techniques for a general assignment problem and for the multitarget tracking problem.

Holographic data are acquired during hydrodynamic experiments at the Pegasus Pulsed Power Facility at the Los Alamos National Laboratory. These experiments produce a fine spray of fast-moving particles. Snapshots of the spray are captured using in-line Fraunhofer holographic techniques. Roughly one cubic centimeter is recorded by the hologram. Minimum detectable particle size in the data extends down to 2 microns. In a holography reconstruction system, a laser illuminates the hologram as it rests in a three- axis actuator, recreating the snapshot of the experiment. A computer guides the actuators through an orderly sequence programmed by the user. At selected intervals, slices of this volume are captured and digitized with a CCD camera. Intermittent on-line processing of the image data and computer control of the camera functions optimizes statistics of the acquired image data for off-line processing. Tens of thousands of individual data frames (30 to 40 gigabytes of data) are required to recreate a digital representation of the snapshot. Throughput of the reduction system is 550 megabytes per hour (MB/hr). Objects and associated features from the data are subsequently extracted during off-line processing. Discrimination and correlation tests reject noise, eliminate multiple-counting of particles, and build an error model to estimate performance. Objects surviving these tests are classified as particles. The particle distributions are derived from the data base formed by these particles, their locations and features. Throughput of the off-line processing exceeds 500 MB/hr. This paper describes the reduction system, outlines the off-line processing procedure, summarizes the discrimination and correlation tests, and reports numerical results for a sample data set.

Small targets defined in the paper are low-RCS targets or weak infrared emitting targets which are very difficult to track by a single sensor (a radar or an IRST) because the measurements gained by the sensor are usually unsuccessive. Although a single sensor can only observe the target in unsuccessive way, the sensor net may sometimes track the target continuously by data fusion of the fractional trajectories or hits of the target from all sensors of the netted system. Typically, data fusion method of a distributed sensor net is track-to-track fusion that requires continuous track- level data (space vector estimates and their covariance matrices) from sensors in the net. For the case of small targets this condition can not be satisfied. In the paper, we derive a method to fuse the data of the small target from the system. The method includes data association and tracking. Special attention is paid to track initiation with Hough transform for the association algorithm. Test and results are also given in the paper.

Two different vehicles which have the same shapes and sizes fly together in upper space at the same velocity. The vehicles are so far away from the sensors that the acquired images are point targets. No shape information can be gotten. These two vehicles fly together and no motion characteristics may be used. Therefore the infrared (IR) radiation characteristics of them are important for discerning these two vehicles. In this paper, three ground-based IR sensors are used to get the IR radiation spectrum of the point targets, and twelve IR characteristics are selected for recognizing them. First, a BP network is used to recognize the point targets at each base. Then a subjective Bayesian method is adopted to fuse the recognized results given by BP networks on three bases at the same time. And the result given by Bayesian is fused by D-S evidence theory with the result at next time till the belief function is more than threshold. The emulation shows that the last outputs is satisfactory.

A new detecting algorithm for close region, which has rectangle or parallelogram shape (e.g. airfield track), is developed in this paper. The traditional detecting algorithms for this kind of close region have some faults: long operating time and high demand for image quality. Mainly based on gray level shifting characteristic and size characteristic of close region, this proposed detecting algorithm for close region is not needed to extract line and search parallel lines. In this algorithm, a new terminology-edge dot couple, which is used as parameter of close region detection, is expressed and the method to extract it is described. A new technology-region growing, which use edge dot couple as growing base, is developed to search out close region. The author applies this algorithm to detect airfield track. The experiment result shows that this method is good for the close region detection.

This paper deals with the problems of point target detection and tracking initialization in infrared image sequence. A processing method based on median filter is adopted to process the original images and remove the correlated spatial noise. An accumulation scheme is implemented to suppress the temporal noise and improve the SNR. Because of the movement of target, it may form a point, a tiny blob or a trace on the accumulated image. The algorithm deals with all of these cases, a segmentation method based on the sum of the gray within the target region is proposed, because the sum is equal for all these cases statistically. Simultaneously the parameters of track initialization can be estimated through the trace of the target. Two infrared image sequence is used to testify this algorithm, the results of simulation indicate that this algorithm reaches a high performance.

A multifocal matrix method for super Rayleigh resolution imaging and determining geometric and dynamic parameters of objects is developed and studied. The key element of the method is a multifocal matrix consisting of focusing elements. The coherent light scattered by an object is focused by these elements and focused fields are detected at elements focused. The results of detection are used for determining the angular coordinates, amplitudes and phases of light fields scattered by different parts of the object surface. If the object under investigation consists of closely spaced small targets that cannot be resolved using the Rayleigh criterion, the method provides a useful tool for determining the angular coordinates, velocity, scattering coefficient and the distance of each target. The effect of the additive noise of focusing detector elements on the angular resolution of the method is analyzed. The method provides two ways of obtaining information about small targets. The first allows the building of their two-dimensional images, the other is for determining their overall dimensions, rotational speeds, parameters of surface roughness.

A new measurement device is presented. Extreme design care is devoted to achieve a closer relation between the structure of the mathematical description and the finite-resolution properties of physical detectors that characterize any real measurement process, than previous attempts described in scientific and technical literature. In fact, experimental data are captured by means of a pair of active sensors placed at a relative fixed distance, working in a coupled arrangement. The presented device domain sensor operates in a discrete variable domain. Beyond the operational advantages in terms of simplicity and computational speed, it agrees with the results of biological observation which reveals that highly structured data always come from coupled transducing bio-elements. A numerical example is presented: an optimal computational precision level can be selected according to the required output precision. Furthermore, it can be shown that for that required precision, the output discrete data set is consistent under inversion transformation.

This paper describes the algorithms that Arete with support from the Office of Naval Research (ONR) is developing for shipboard infrared search and track (SIRST) detection of low-observable targets, such as subsonic, sea-skimming cruise missiles. Early detection of low signal-to-noise (SNR) targets (6 - 10 dB) is provided by Arete's Bayesian field tracker (BFT), which is a track- before-detect algorithm. Candidate detections from the BFT are used to initialize the position, velocity and likelihood of candidate tracks in a peak likelihood track-after-detect tracker. False alarm mitigation is accomplished in part by requiring the temporal evolution of a candidate track to be consistent with that of an incoming sea-skimming cruise missile. The overview of the algorithms involved in these trackers and results from both real targets and simulated targets injected into real images are presented.

For track maintenance, there are primarily three generic sensor data fusion algorithm architectures, namely, central fusion, track fusion, and what will be referred to as composite measurement fusion. In central fusion, the sensor measurements are distributed by each sensor and the measurements from multiple sensors are then used to update the global tracks. In contrast, in track fusion a sequence of measurements is processed at the sensor or platform level to form tracks that are then distributed and this track data is then used to update the global tracks. Track fusion is also sometimes called hierarchical fusion, federated fusion or distributed fusion. Finally, in composite measurement fusion the measurements from multiple sensors for each apparent target are first combined to form a composite measurement and then the composite measurements are then used to update the global tracks. In addition there is also hybrid fusion, which is a combination of at least two of the above fusion approaches. The tracks typically include features or other target classification information. Each of these algorithm architectures has their own advantages and disadvantages. For example, track fusion may lead to substantially reduced communications load and that can be very important for physically distributed platforms. Track fusion also tends to be less sensitive to residual sensor bias errors. Central fusion typically provides more timely information. Also, for certain types of target scenarios and sensor suites, central fusion provides better accuracy in both estimation and target classification. Recent developments in track fusion make a particular hybrid fusion algorithm architecture not only appealing but practical. In this hybrid fusion, either measurement or track data in the form of a tracklet is distributed from a sensor for a target. This approach offers many of the advantages of both the central and track fusion algorithm architectures. Recently a number of methods have been identified for track fusion that take into account the cross- correlation between tracks from different sensors for a target. A tracklet, for example, is track data for a target that is not cross-correlated with any other track data for that target. Hence a special case of track fusion is tracklet fusion. Many of these track fusion methods do not, however, address the cross-correlation caused by process noise. Also, in track fusion, the selection of the mathematical model used for process noise is more critical than is commonly thought. This paper addresses these two issues.