A morphology-based algorithm has been developed for point target detection in IRST applications. It exhibits comparable detection and false alarm performance to a median filter. The morphology-based algorithm has an efficient computational paradigm based on combinations of simple nonlinear grayscale operations, which makes it ideally suited to real- time, high data rate IRST applications. A detection filter based on morphological background estimation exhibits spatial high-pass characteristics emphasizing target-like peaks in the data and suppressing all other clutter. Example cases are presented which point out the detection performance differences between the morphological and median approaches. Overall performance results were generated in the form of ROC curves for cloud, horizon and sea clutter IRAMMP backgrounds.

This paper presents a comparative performance evaluation of three signal processing architectures for the point-target detection problem in infrared surveillance systems. The comparison involves an adaptive linear spatial filter of the least-mean-square (LMS) type, with a median subtraction filter, and a morphological based processor. The median and morphological filters are nonlinear operators that are based on the order-statistics of the input samples. The three architectures are exercised, via computer simulation, against real deterministic data sets comprising a range of horizon backgrounds. Receiver operating characteristic (ROC) curves are presented and provide a quantitative measure for the performance comparison. Some advantages and disadvantages of each filter type are indicated, with recommendations for a hybrid architecture which incorporates the different filtering schemes.

The requirements for high detection probability and low false alarm probability in modern wide area surveillance radars are rarely met due to spatial variations in clutter characteristics. Many filtering and CFAR detection algorithms have been developed to effectively deal with these variations; however, any single algorithm is likely to exhibit excessive false alarms and intolerably low detection probabilities in a dynamically changing environment. A great deal of research has led to advances in the state of the art in Artificial Intelligence (AI) and numerous areas have been identified for application to radar signal processing. The approach suggested here, discussed in a patent application submitted by the authors, is to intelligently select the filtering and CFAR detection algorithms being executed at any given time, based upon the observed characteristics of the interference environment. This approach requires sensing the environment, employing the most suitable algorithms, and applying an appropriate multiple algorithm fusion scheme or consensus algorithm to produce a global detection decision.

Advanced radar receivers having an instantaneous dynamic range exceeding 90 dB are currently under development. These receivers must operate linearly to achieve satisfactory performance in a severe clutter environment. Presently available, state-of-the-art analog-to- digital converters (ADCs) having sufficiently high sampling rates do not have adequate dynamic ranges to digitize the outputs of these receivers. A technique has been developed that effectively reduces the dynamic range of the input signal prior to digitization. The dynamic range reduction is accomplished through a process that predicts the next radar return signal from the previous return signals, generates a replica waveform, and subtracts this replica waveform from the radar return signal prior to digitization. This process allows the radar return signal to be digitized without distortion by an ADC having a limited dynamic range. The full dynamic range of the radar return signal is then restored by adding the replica waveform to the ADC output. Tests and evaluations using both synthetic and recorded radar data have demonstrated in excess of a 30 dB reduction in the dynamic range of the signal at the ADC input when strong clutter is present.

In this paper we discuss the estimation of separation and polarization characteristics for a two point target from interferogram when laser backscattered radiation is collected with an incoherent array of polarization sensitive detectors. We establish fundamental limits on the resolution of such a system and obtain a quantitative understanding of the dependence of estimation accuracy on target separation, polarization state and number of photons available. The results obtained in semiclassical limit include Poisson photon statistics. The performance of various estimators is evaluated by utilizing the Cramer-Rao bound, for both signal and background shot noise limited conditions. The results presented are applicable to laser radars, discrimination of closely spaced objects, and provide a better understanding of the system when spatial polarization coding of source is used to transmit information.

There has been a growing interest in employing infrared (IR) detectors to locate and track multiple sources in recent years. Conventional methods of locating point sources such as centroiding are not effective when point sources are closely spaced. In this paper, alternative location finding methods with the potential of resolving closely spaced objects (CSOs) is introduced. Of the three algorithms introduced here, two are based on the eigendecomposition of the input data. The other is predicated on least squares error modeling (LSE) with a Gram- Schmidt orthogonalization step to ensure fast convergence. Resolution capabilities of these algorithms are compared through Monte Carlo simulations at various noise levels. Estimates obtained through the LSE modeling approached the Cramer-Rao lower bound for high signal- to-noise-ratios. The performance of the LSE estimate is severely degraded in the presence of nongaussian noise. An outlier detection scheme that may be used in conjunction with the location and amplitude estimation procedure is described. Its effectiveness is demonstrated through Monte Carlo simulations.

A technique is described for recovering positional and radiometric information on unresolved objects that are so closely spaced that their individual blur functions overlap. Emphasis is on point sources. A Bayesian Spectral Analysis method has been modified to two dimensions and applied to resolving 'clumps' of objects for both simulated and real data. The method is able to judge the amount of noise in the data and provide error bars in the individual pulse positions and amplitudes from a single data set rather than from the deviations observed after measuring many independent sets of data. The Bayesian technique can also estimate the number of discrete objects in a given clump. Noisy simulated data containing three sources was fitted by a one-, two-, three-, and four- source model. By the way it formulates the model, the Bayesian approach naturally includes a factor which reflects the reduction in the number of degrees of freedom for a model with a greater number of sources. As a result, the algorithm gives a higher probability for the three-source model than for the four-source model while resoundingly rejecting the one- and two-source models. The estimated centroids and amplitudes are shown to agree with the truth within the derived error bars to the degree expected by gaussian errors. Studies of data taken during a flight test by a sensor that measured a scene simultaneously in the visible and long-wavelength regions show that positional information derived from visible-wavelength data can be 'fused' with infrared images to derive the LWIR intensities of individual objects in a unresolved clump. The estimated LWIR intensities using the visible assist are shown to be an improvement over working with the LWIR data alone.

Closely spaced object (CSO) processing methods often utilize a low pass filter plus deconvolution process for resolution enhancement. Other methods utilize an iterative least squares fitting process employing either a prior knowledge for initial estimates, or parameterizing initial estimates until an optimal fit is achieved. Each method has advantages and disadvantages. Low pass filtering plus deconvolution, performed in the time domain with a finite impulse response filter (FIR), is less computationally intensive than nonlinear least squares fitting. Nonlinear least squares fitting is generally superior as a maximum likelihood estimator of amplitude and position for lower signal to noise ratio (SNR) CSO clusters. Here we present a method of combining the two processes by using cues derived from deconvolved data as initial estimates for least squares fitting. The process is demonstrated on raw sensor data from the Airborne Surveillance Testbed (AST) sensor. Effects of apparent focal plane motion on CSO resolution enhancement and least squares refinement are discussed. Several least squares refinement methods are presented.

As the sensitivity of next generation IRST systems is improved and the need for detecting dim targets becomes important, it becomes of paramount importance to mitigate the false alarms which occur due to clutter artifacts so that noise limited performance can be achieved. One of the chief sources of false alarms in current IRST systems operating in cloud and ground clutter are scenes with high edge content. Clutter with acute angles has a large amount of power in the mid-band spatial frequencies and will compete with target energy out of the matched filter. Since the simplistic approach of just blanking edge regions would cause targets to be lost, a more sophisticated procedure needs to be developed. The method of false alarm mitigation (FAM) developed in this study is to construct a likelihood ratio test by modeling the probability density of the local SNR discriminant as a combination of a Gaussian and a Gamma distribution and by modeling the edge discriminant with a central chi density when no edge is present and a noncentral chi density when an edge is present. The output of the likelihood ratio test leads to a decision region in the two-dimensional discriminant space for deciding when a target is present versus when a target is absent.

When the thermal contrast between a target and the background is large it is usually sufficient to use a single spectral band for detection purposes and consequently spectral emissivity variations are ignored. We consider the problem of detection and estimation of spectral emissivity in the thermal infrared emphasizing the particular detection situation where the target thermal contrast is small. More generally, we show that temperature can be considered as random interference of low rank. This leads to a procedure of optimally estimating both the temperature and spectral emissivity components. of radiance measurements. In addition to temperature and spectral emissivity estimation we outline a phenomenology based optimal detector that is invariant to unknown levels of temperature variation and unknown levels of target thermal contrast.

This paper addresses pre-detection signal processing of dual-band detection algorithms, a generalized likelihood ratio test and modified sample matrix inversion, are developed for the case of gaussian interference. The associated theoretical detection performance and computational requirements are also derived. General detection performance formulas are provided which can be used to determine the conditions for which the dual-band algorithms provide the greatest improvement in performance. The utility of the dual-band detection algorithms is demonstrated by comparison with the optimal single- and dual-band likelihood ratio tests as well as the single-band GLRT and MSMI algorithms. The dual-band algorithms are also compared with a system using two single-band filter channels whose outputs are combined after thresholding. Results are generated from theoretical formulas and from a simulation, developed at GE, which duplicates the filtering and detection functions of the algorithms.

This paper will present a new algorithm for determining the presence or absence of a weak signal in non-Gaussian noise. A weak signal is one that is vanishingly small compared to the noise disturbance. There are several applications in which detecting a weak signal is important. If the weak signal is the result of the reflection of a small target then accurate detection can indicate target presence. If the target is maneuvering and measurements can only be made at widely separated fixed intervals, then after target detection an estimate of target velocity can be made. The algorithm presented in this paper is no more structurally complex than the LOD, yet possesses several important advantages over the LOD. The fundamental advantage is the fact that the underlying noise statistics do not have to be known a prior. In addition, whereas the LOD may require a rather complex nonlinearity to preprocess the data, the algorithm presented here does not. This paper will develop the algorithm, and then report on simulation testing that was performed to ascertain its performance. It will be shown that the proposed algorithm performs significantly better (that is, is able to detect the presence of weak targets) than more conventional linear detection methods (Wiener filtering).

In the signal detection problem in nonstationary cluttered backgrounds using data from an imaging sensor, the receiver structure must generally be adapted to the local clutter statistics. Typically, the receive:r is implemented as a linear matched filter and the local adaptation consists of estimating and inverting a local covariance matrix to obtain the optimum weight vector. The local estimates of the inverse covariance matrix can be obtained in a variety of ways, but is generally a computationally expensive procedure and is prone to inaccuracies whenever estimation windows overlap clutter region boundaries. In the present paper we describe a simple but effective adaptive detection procedure which avoids some of the difficulties associated with existing schemes. This procedure employs a simple least mean-square (LMS) algorithm to adapt a linear matched filter to maximize local SNR. We describe a particular multiscan version of this algorithm with improved convergence properties. In particular, by implementing multiple parallel scanning patterns it's possible to avoid potential convergence problems at region boundaries associated with conventional single-scan adaptive approaches. Finally, we describe the performance of this scheme and compare its performance with competing approaches. We demonstrate performance approaching that achievable when perfect knowledge of local clutter statistics are available.

This paper presents a new, self-adaptive technique for the correlation of non-uniformities (fixed-pattern noise) in high-density infrared focal-plane detector arrays. We have developed a new approach to non-uniformity correction in which we use multiple image frames of the scene itself, and take advantage of the aim-point wander caused by jitter, residual tracking errors, or deliberately induced motion. Such wander causes each detector in the array to view multiple scene elements, and each scene element to be viewed by multiple detectors. It is therefore possible to formulate (and solve) a set of simultaneous equations from which correction parameters can be computed for the detectors. We have tested our approach with actual images collected by the ARPA-sponsored MUSIC infrared sensor. For these tests we employed a 60-frame (0.75-second) sequence of terrain images for which an out-of-date calibration was deliberately used. The sensor was aimed at a point on the ground via an operator-assisted tracking system having a maximum aim point wander on the order of ten pixels. With these data, we were able to improve the calibration accuracy by a factor of approximately 100.

We propose a technique to obtain directly from a series of photon-limited frames, the spectrum of an object moving with constant velocity. Instead os averaging statistical functions as the auto or triple correlations or their Fourier transforms, our method averages the series of frame spectra once the phase factor due to the movement has been removed. Two different procedures to obtain this phase factor are studied: the temporal derivative of the spectrum logarithm and the temporal Fourier transform of the series of spatial spectra. The latter method involves a larger number of calculations but it produces much better results, especially when only a small number of frames are available. Finally, the recovery technique is checked for a simulated experiment in which a one-dimensional object is reconstructed from a short series of photon-limited frames.

Several factors are important in the development and evaluation of signal and data processing algorithms. Critical among these factors is the accurate representation of the backgrounds against which targets may need to be detected and tracked. A background model has been developed in the Advanced Surveillance Testbed which supports the evaluation of multi-target algorithms associated with electro-optical sensors. In this paper, a discussion of the synthetic background model which is part of the Advanced Surveillance Testbed, its validation and its utility is given. The synthetic background model employs an equation-based design and, as it currently exists, is composed of a cloud model, a terrestrial model and radiative transfer calculations. The validation of the synthetic background model has been done against both theoretical results and full background scenes; results are presented and discussed. The utility of the equation-based design in providing flexibility, convenience and capability is presented and discussed.

A real-time multiresolutional approach for target tracking is developed in this paper. The wavelet transform is utilized to provide the multiresolutional measurements and bridge information at different resolutional levels. The approach is applied to tracking maneuvering targets and novel results are obtained.

A sequential hypothesis testing approach is proposed for multi-frame detection and tracking of low-observable, maneuvering point-source targets in a digital image sequence. The resultant Multiple Multistage Hypothesis Test Tracking (MMHTT) algorithm extends tracks formed from sequentially-detected target trajectory segments using a multiple hypothesis tracking strategy. Computational efficiency is achieved by using a truncated sequential probability ratio test (SPRT) to prune a dense tree of candidate target trajectories and score the detected trajectory segments. Results of an analytical evaluation of the algorithm's performance are discussed in relation to experimental results from an optical satellite tracking application.

In our previous work we presented a method for precision tracking of a low observable target based on data obtained from imaging sensors. The image was divided into several layers of gray level intensities and thresholded. A binary image was obtained and grouped into clusters using image segmentation techniques. Using the centroid measurements of the clusters, the Probabilistic Data Association Filter (PDAF) was employed for tracking the centroid of the target. In this paper, the division of the image into several layers of gray level intensities is optimized by minimizing the Bayes risk. This optimal layering of the image has the following properties: (a) following the segmentation, a closed-form analytical expression is obtained for the single frame based centroid measurement noise variance; (b) in comparison to the measurement noise variance is smaller by at least a factor of 2, thus improving the performance of the tracker. The simulation results presented validate both the expression for the measurement noise variance as well as the performance predictions of the proposed tracking method.

A real time system with high speed hardware and efficient algorithms has been developed. This system is applied to small targets detection in IR image. An adaptive LMS filter is used to eliminate wideband background. Interframe analysis rejects the noises and point-like clutters. Parallel processing makes it possible to obtain features of targets every frame. The experiment result shows that the accuracy of target position is within half pixel.

The space track catalog of satellites in orbit is generally maintained using analytic methods. New technology developments in the area of parallel processing provide the capability to apply more exact methods to determine and maintain the ephemerides of a larger number of space objects more precisely. Space object tracking accuracy is becoming increasingly important in space programs such as the Space Station, where the collision hazard is critical, and for military application requiring precise positioning of satellites. Affordable massively parallel processing architectures will soon be available to address this problem. We expect that before long the number of processing elements in a single affordable box will approach or exceed the number of satellites in orbit. In this paper we consider algorithms and architectures for processing a large number of space objects in a parallel sense. These improvements will enable the tracking of many small objects with precision and will improve the confidence with which collision hazards can be assessed. In addition, as sensor capabilities are improved through technology upgrades, the accuracy of these computational methods will continue to exceed the precision of the measurements.

Tracking algorithms commonly use practical models of target motion to estimate the target's kinematic quantities such as the position, the velocity and in certain cases, the acceleration. When there is a maneuver, the tracking algorithm should detect the error created by this change and correct the situation to adapt itself to this new change or new tracking model. There are different approaches in the literature for handling maneuver detection using different filtering techniques. A thorough literature survey about different types of filtering techniques used for maneuver detection has been performed. The focus of this study has been the parallel filtering techniques. Some of those techniques given by different authors are summarized in this paper. This paper presents a parallel filter design using three linear Kalman filters with a simple switching algorithm for maneuver detection selected for the Multi Sensor Data Fusion (MSDF) for an anti-air warfare (AAW) surveillance radar. This design is relatively simple compared to other parallel Kalman filter techniques and requires modest computer resources. The parallel filter design has been compared with a single Kalman filter design previously used. The simulation results have shown a great deal of improvement with parallel filtering, particularly in speed estimations and in filtering stability when a target is maneuvering.

Analytic expressions for allowable number of lost track updates are derived based on tracker characteristics that determine the track deletion logic for maneuvering and nonmaneuvering targets. The tracker characteristics chosen are: filter gain, track update time, measurement correlation gate size, target maneuver and measurement error variance.

The crucial problem in multiple target tracking is the hit-to-track data association. A hit is a received signal from a target or background clutter which provides positional information. If an incorrect hit is associated with a track, that track could diverge and prematurely terminate or cause other tracks to also diverge. Most methods for hit-to-track data association fall into two categories: Multiple Hypothesis Tracking (MHT) and Joint Probabilistic Data Association (JPDA). Versions of MHT use all or some reasonable hits to update a track and delay the decision on which hit was correct. JPDA uses a weighted sum of the reasonable hits to update a track. These weights are the probability that the hit originated from the target in track. The computational load for the joint probabilities increase exponentially as the number of targets increases and therefore, is not an attractive algorithm when expecting to track many targets. This paper reviews the JPDA filter and two simple approximations of the joint probabilities which increase linearly in computational load as the number of targets increase. Then a new class of near optimal JPDA algorithms is introduced which run in polynomial time. The power of the polynomial is an input to the algorithm. This algorithm bridges the gap in computational load and accuracy between the very fast simple approximations and the efficient optimal algorithms.

This paper discusses probabilistic, possibilistic, and symbolic logic approaches to data association as part of the multi-sensor integration (MSI) problem. Alternative probabilistic scoring equations are given with specific trades between max a posterori and chi-square scoring. Dempster-Shafer and fuzzy set approaches to treat data with high uncertainty-in-the- uncertainty are compared. The role of symbolic rule-oriented approaches especially as part of a fusion analyst workstation (FAWS) expert system is discussed. Lastly, an avionics MSI design trade example is presented.

This paper presents a knowledge/rule base system which controls parameters and modes for multi-function/multi-mode sensors such as radar. The system performs a situation assessment of the signal/noise environment followed by appropriately prioritized automatic control of the parameters and modes of the sensor system. The work involved software development of an advanced simulated radar interacting with a simulated environment. The most recent activities involved the implementation of the system with a phased array sensor in a real world signal/noise environment. The results illustrate the improvement in signal to noise as the system assesses the noisy sensor environment and consequently controls the sensor's parameters and modes. A monitor display also reports the type of noise assessed by the system, i.e. interference, terrain clutter, weather clutter, etc., along with the course of action taken, i.e. sidelobe blanking, frequency agility, Doppler filtering, etc., to suppress such noise.

The target tracking case, where radar and IRST contacts are fused at the central level fusion architecture, with special attention to the coordinate system, has been analyzed in this presentation. Tracking targets by fusing contacts from dissimilar sensors in a central level fusion process is acknowledged to be the most powerful tracking technique. This approach maximizes the synergy among the data, consequently the accuracy of the tracking. Among the various tactical possibilities offered by data fusion, this approach permits to perform accurate tracking when the platform is 'almost' electromagnetically silent. Tracking targets with sparse use of radar is possible only if the angle contacts obtained from passive sensors are fused optimally with the parsimonious and intermittent radar contact. Usually, the range and angle data provided by a surveillance radar permit to use cartesian coordinates and track with a conventional Kalman filter using linear target model and measurement model equations. On the other hand the nature of the angle-only data provided by a passive sensor does not permit to take advantage of the linear models and the IRST tracker provides ambiguous angle tracks. To benefit from central level fusion, the state estimation process in the fusion function must be performed in a coordinate system that accommodates the radar as well as the IRST contacts. A simple simulation is done in three dimensions and in three different coordinate systems: cartesian, spherical and modified spherical. The analysis shows the feasibility of data fusion in all three coordinate systems.

An algorithm is presented to relatively align a 3D sensor and a 2D sensor measuring range and azimuth. The alignment is done using common targets that are tracked by both sensors. The algorithm estimates and removes sensor biases and sensor frame orientation errors. For illustrative purposes, the alignment algorithm is applies to simulated track data.

For multisensor fusion systems it is a prerequisite to accurately estimate and correct all systematic errors. Adequate estimation methods only exist if all systematic errors are constant random variables, while in practice they may change with time. When the object states, the systematic errors and the observations vary according to a linear Gaussian system, then one large Kalman filter forms the optimal estimator for the combined state of all object states and all systematic errors. In general the numerical complexity of this Kalman filter prohibits practical application. In order to improve this situation we decouple the large Kalman filter into a number of separate filters: for each object one track maintenance Kalman filter, and for the estimation of all sensor related systematic errors one Kalman-like filter, which we call the Macro filter. The effectiveness of this approach is illustrated through simulations for a simple example.

One of the more difficult targets to track is an aircraft performing high speed maneuvers. The Interacting Multiple Model (IMM) algorithm uses multiple models that interact through state mixing to track a target maneuvering through an arbitrary trajectory. The IMM algorithm provides significantly better tracking results when compared to a single model track filter. To improve tracking performance, multiple sensors can be used to provide more information about the target. Using the measurements from several sensors with a single motion model track filter can provide improved track performance when compared to a single sensor system. Since single sensor track filters use decision-directed approaches for maneuver response, using multiple sensors with a single model track filter would be difficult to implement because periodic measurement updates cannot be expected and the sensors may be dissimilar with different accuracies. Thus a very complex tracking algorithm would be required. While the tracking performance of a single model may improve with additional sensors, it can be erratic for maneuvering targets. Using multiple sensors with the IMM algorithm can improve the IMM algorithm performance without the erratic performance exhibited by single model trackers. Target tracking with multisensor systems is described along with the IMM algorithm. Comparisons of track performance with IMM algorithm and single motion model track filters are presented for several sensor systems.

The SME filter approach to multiple target tracking (MTT) is extended to the case when position measurements are provided by more than one sensor. In this approach to MTT, there is no need to consider target/measurement associations or sensor-to-sensor correlations. The SME filter is generated for the case of N targets and M sensors, with the target motions assumed to consist of random perturbations about constant-velocity trajectories. The key idea of the SME approach is to form a new measurement vector by taking sums-of-products of the Cartesian coordinate measurements generated from spherical coordinate measurements provided by each sensor. The new measurement vectors at each sensor are either 'summed' or 'stacked' to form a composite measurement vector in each coordinate. This results in a 'sum SME filter' and a 'stack SME filter'. A computer simulation is given which shows that the stack SME filter can result in much better performance than that of the sum SME filter.

Several multiple model techniques have been applied to the tracking of maneuvering targets. The two techniques which provide the best tracking performance for maneuvering targets are the Second Order General Pseudo-Bayesian (GPB2) and Interacting Multiple Model (IMM) algorithms. In both algorithms, the dynamics of the system is represented by multiple models which are hypothesized to be correct and model switching probabilities governed by a first order Markov process. The authors have developed an extension of the IMM algorithm, the second order Interacting Multiple Model (IMM) algorithm, which provides improved tracking performance when compared to that of the IMM and GPB2 algorithms for applications with large measurement errors and low data rates. In the IMM2 algorithm, the state estimate is computed under each possible model hypothesis for the two most recent sample periods with each hypothesis using a different combination of the previous model- conditional estimates. Thus, the IMM2 algorithm requires r² filters for r models. The development of the IMM2 algorithm is given along with a summary of multiple model estimation for tracking maneuvering targets and simulation results for the IMM, GPB2, and IMM2 algorithms.

This paper presents an analytical performance model, verified by Monte Carlo simulation, that predicts the track acquisition performance of a simple, single sensor tracking algorithm. The tracking algorithm consists of a constant velocity Kalman filter, a pure splitting data association algorithm and an M/N/K promotion algorithm. The M/N/K promotion algorithm declares a track to be firm if it receives M hits within N frames of track start and deletes a firm track if it receives K misses in a row. Track acquisition is evaluated in terms of the expected number of frames to declare a firm track on a real target and the spatial density of false tracks that are declared to be firm. The expected time to firm track is derived by calculating the first passage time for the Markov chain modeling target track acquisition, and false track generation is modeled as a multi-type branching process.

A multiple hypothesis tracking (MHT) system produces multiple data association hypotheses that potentially consist of several tracks on the same target. The relative likelihood of these tracks changes as data are received. Also, the manner in which an MHT system produces and deletes tracks leads to a potential discontinuity in the track numbers associated with a given target. Thus, the direct output of an MHT tracker may be difficult to interpret on a system display or to use to perform track-to-track association in a multiple sensor tracking system. This paper presents a methodology so that the best tracks on a set of targets can be identified and a link provided with previously identified tracks. This allows a multiple sensor tracking system to maintain track-to-track association histories over time. Also, although track numbers change internal to the MHT logic, the methods presented in this paper lead to a user output such that a single track number is maintained on a given target throughout an encounter. Thus, this paper presents a solution to one of the major practical problems associated with the implementation of an MHT system.

The ever-increasing demand in surveillance is to produce highly accurate target and track identification and estimation in real-time, even for dense target scenarios and in regions of high track contention. The use of multiple sensor, through more varied information, has the potential to greatly enhance target identification and state estimation. For multitarget tracking, the processing of multiple scans all at once yields high track identification. However, to achieve this accurate state estimation and track identification, one must solve an NP-hard data association problem of partitioning observations into tracks and false alarms in real-time. The primary objective in this work is to formulate a general class of these data association problems as a multidimensional assignment problem to which new, fast, near-optimal, Lagrangian relaxation based algorithms are applicable. The dimension of the formulated assignment problem corresponds to the number of data sets, and the constraints define a feasible partition of the data sets. The linear objective function is developed from Bayesian estimation and is the negative log likelihood function, so that the optimal solution yields the maximum likelihood estimate. After formulating this general class of problems, the equivalence between solving data association problems by these multidimensional assignment problems and by the currently most popular method of multiple hypothesis tracking is established. Track initiation and track maintenance using an N-scan sliding window are then used as illustrations.

The symmetric measurement equation (SME) filter for track maintenance in multiple target tracking is extended to the general case when there are an arbitrary unknown number of false and missing position measurements in the measurement set at any time point. It is assumed that the number N of targets is known a priori and that the target motions consist of random perturbations of constant-velocity trajectories. The key idea in the paper is to generate a new measurement vector from sums-of-products of the elements of 'feasible' N-element data vectors that pass a thresholding operation in the sums-of-products framework. Via this construction, the data association problem is completely avoided, and in addition, there is no need to identify which target measurements may correspond to false returns or which target measurements may be missing. A computer simulation of SME filter performance is given, including a comparison with the associated filter (a benchmark) and the joint probabilistic data association (JPDA) filter.

Kamen et al. have developed the symmetric measurement equation (SME) filter as an alternative to multi-target trackers based on data association. This paper presents an improved multi-dimensional SME tracking algorithm which agrees with Kamen's for one-dimensional scenarios and avoids the ghost target problem in higher dimensions. In addition, we provide a more efficient method for computing the noise covariance matrix of the SME coefficients. This was the major computational bottleneck of earlier SME implementation, and we have reduced its complexity from at least 2^N/2 operations to at most D⁴N⁵, where N is the number of targets and D the number of dimensions. Computer simulations illustrate a failure mode that the new algorithm avoids, and gives a sample comparison to a standard data- association algorithm, global nearest neighbor.

This paper deals with global track initiation problem and our purpose is to apply a quite efficient method canned the Multiple Hypothesis Filter to solve it. This original measurement- oriented approach in its optimal version leads to an ever-expanding tree and to a computational problem so that some reduction procedures have to be applied. We first present a state-of-the- art of methods that propose solutions to the hypothesis control problem. Then, we introduce a new criterion in order to control pruning action and avoid losing important information.

This paper describes how optical sensor signal processing and data association methods that have been developed for Aerospace applications can be applied to the traffic monitoring function of Advanced Traffic Management Systems (ATMS). It first discusses techniques that have been developed for background estimation and detection of vehicles on a roadway. Then, the transformation to tracking coordinates and the multiple target tracking (MTT) algorithm that produces traffic flow observation data are outlined. An extended Kalman filter that takes observed flow data from multiple sensor sites and produces flow estimates for an entire roadway is described and its application to incident detection discussed. Preliminary results using simulated and actual freeway data are presented. Finally, techniques for presenting this data to the user and the manner in which these signal and data processing techniques relate to an overall ATMS design are outlined.

The central problem in multitarget-multisensor tracking is the data association problem of partitioning the observations into tracks and false alarms so that an accurate estimate of the true tracks can be recovered. Many previous and current methodologies are based on single scan processing, which is real-time, but often leads to a large number of partial and incorrect assignments, and thus incorrect track identification. The fundamental difficulty is that data association decisions once made are irrevocable. Deferred logic methods such as multiple hypothesis tracking allow correction of these misassociations and are thus considered to be the method for tracking a large number of targets. The corresponding data association problems are however NP-hard and must be solved in real-time. Such algorithms have been developed in earlier work of the authors and the intent of this work is to demonstrate the efficiency and robustness on a class of tracking problems.

Since phased array radars have the ability to perform adaptive sampling by the radar beam, proper control of the radar has the potential for significantly improving many aspects associated with the tracking of multiple maneuvering targets. The technique proposed in this paper uses the Interacting Multiple Model (IMM) algorithm to track maneuvering targets and control the sampling time and energy levels. Since the output of the IMM algorithm better represents the accuracy of the state estimates during a maneuver than a single model filter, the IMM algorithm is used to compute and on-line measure of tracking performance to determine the scheduling time of the next track update sample period in order to maintain a given level of performance. The sample time is computed as the one positive root of a polynomial equation of the sample period. The model probabilities of the IMM algorithm are also used to schedule the energy level of a radar dwell. As a result, the update times for the filter are a function of track filter performance and the target trajectory. Algorithms for computing the sample time and energy level using the output of the IMM algorithm are developed in this paper. Performance comparisons are given for the IMM algorithm using constant data rates, scheduled energy levels, and adaptive data rates.

QUICKTRACK is a new track initiation algorithm for optical sensors tracking ballistic targets. It is designed to compute new tracks at the earliest moment mathematically possible to any desired accuracy up to the limiting accuracy of the data themselves. Initial testing now underway shows very promising results. Whether computational requirements in realistic cases will be lower or higher than with conventional methods is yet to be determined. QUICKTRACK can initiate confirmed 3 dimensional (3D) tracks of targets viewed by 1, 2, 3, or 4 sensors although it offers no special advantages for those targets viewed by only 1 target. In a particular case discussed where conventional track initiation needed 200 seconds of data, QUICKTRACK initiated track after only 60 seconds of data.

A critical problem in current missile guidance systems is that of tracking targets based on visible light or infrared imaging systems. There is a need to test video tracking system algorithms and hardware without the expense of firing a missile. The work described in this paper, fills the need for low-cost, Personal Computer (PC) based, portable evaluation of tracking systems. This tracker evaluation system consists of a PC, compatible non- developmental hardware, and software to generate real time video and analyze the tracker results. The testing platform can use either a serial (RS232-C) or parallel interface to communicate with the tracker. The PC software is driven by a powerful easy-to-use script language that can be used to select a set of standard tests or describe a fine-tuned customized test to simulate a specific scenario. After the tracker has been tested, the PC software performs statistical analysis on the tracker results to determine tracker performance. PC based closed loop performance evaluation is used to obtain a quick and reliable set of tracker performance values. These values are used to ensure that a tracker meets its specifications, to determine the ability of the tracker to meet the needs of a particular application, and to compare the performance with other trackers.

In this paper, we present a new method for detecting moving dim point targets in image sequences. The algorithm calculates autocorrelation function of projected images including segments of moving target track. The correlation can enhance target intensity and reduce noises. Then, we use Hough transform to determine straight line tracks of targets and to remove false alarm points. The signal-to-noise ratio of images in our experiences is less than 3. Simulation results show that our method can detect the moving targets effectively and is feasible.

Multitarget, multisensor passive sonar tracing has many similarities to the radar problem. In particular, data association issues are critical to overall system performance when tracking barely detectable acoustic sources in a very cluttered ambient background. The purpose of this paper is to provide an overview of some of the crucial data management issues in passive acoustic tracking for those who are familiar with radar tracking. The problem of associating multiple detections in time to form input sequences to nonlinear target trackers is discussed. The problem of associating these input sequences, or contacts, with particular sources and then assigning the contacts to localization and tracking resources will also be described.

A sensor-to-track assignment algorithm is developed for a multisensor tracking system to track multiple maneuvering targets. These sensors are spatially distributed and track multiple targets at the same time. To maintain track continuity of approaching targets is the keynote of this algorithm. When local sensors disagree on which central track to update with their local tracks, this algorithm provides an effectiveness function to resolve the conflict. When measurements from all sensors are used to update central tracks, this algorithm can select or request measurements from the more effective sensors to track potential maneuvering targets.

The study on the tracking of moving target in mm-wave radiative remote sensing is discussed in this paper. First, the estimation equations in disrete state of the target are obtained from the continuous state equations through the analysis of the properties of the moving target and its disturbance noise, then the measurement equation is discussed based on the mm-wave radiative image of the target, and after that the model of target tracking is introduced from the estimation and measurement equations by using Kelman filtering method and the frequency- domain technique in two-dimensional FFT of the mm-wave radiative image. It shows the moving target can be tracked suitably by using this model in mm-wave radiative remote sensing.

Detection and tracking of a moving point target in a low signal-to-clutter plus noise environment is studied in this paper. The clutter plus noise background in sequential images are assumed to be spatially correlated. Two-dimensional adaptive correlation canceling technique is employed to suppress the background noise. An effective searching method for target moving path in image sequence is proposed to find the target position by simple computation as possible.

An effective target tracking system must have the ability to grab and identify various targets automatically. This paper presents a novel algorithm to detect moving targets independent of target's position and shape from image sequence consisted of complex natural scene, and discusses some constraints for removing noises and identifying varied moving targets. For image sequences consisted of tanks moved in an open country, several experimental results (photographs) are given out to show the effectiveness of this algorithm.

Many targets of interest provide only very small signature differences from the clutter background. The ability to detect these small difference targets should be improved by using data which is diverse in space, time, wavelength or some other observable. Target materials often differ from background materials in the variation of their reflectance or emittance with wavelength. A multispectral sensor is therefore considered as a means to improve detection of small signal targets. If this sensor operates in the thermal infrared, it will not need solar illumination and will be useful at night as well as during the day. An understanding of the phenomenology of the spectral properties of materials and an ability to model and simulate target and clutter signatures is needed to understand potential target detection performance from multispectral infrared sensor data. Spectral variations in material emittance are due to vibrational energy transitions in molecular bonds. The spectral emittances of many materials of interest have been measured. Examples are vegetation, soil, construction and road materials, and paints. A multispectral infrared signature model has been developed which includes target and background temperature and emissivity, sky, sun, cloud and background irradiance, multiple reflection effects, path radiance, and atmospheric attenuation. This model can be used to predict multispectral infrared signatures for small signal targets.

The DARPA Multi-Spectral Infrared Camera (MUSIC) sensor system was deployed to Australia in July 1991. A series of measurements on targets and backgrounds were performed over a six week period. Two experiments were performed to investigate the use of multi- spectral techniques to enhance extended targets. The first of these was an experiment to detect ship wakes in the tropical ocean. For this experiment data was taken along a ship wake in both an MWIR band and an LWIR band. The thermal structure of the ocean surface showed a high level of spectral correlation even at the 10 millikelvin sensitivity level of the data. Several ship wakes were detected by processing the data spectrally to remove the background. A second extended target experiment was conducted to detect chemical vapor plumes over a cluttered terrain background. In this experiment, two neighboring infrared bands were chosen, one centered on a chemical absorption band and the second placed away from it. The background clutter proved to be highly correlated between these bands allowing its removal using multi- spectral processing. The enhanced chemical plume was then detected. Results from the processing of the data from both of these experiments will be presented along with a description of the algorithm used.

As detection processing becomes increasingly advanced, for example, in infrared search and track (IRST) systems, the detection threshold becomes the bottleneck to overall system performance. Significantly reducing this threshold requires the capability to track targets in a high clutter environment. In theory, the multiple hypothesis tracking (MHT) algorithm is a solution to this problem. However, in practice, MHT in its basic form becomes computationally prohibitive for all but low to moderate false alarm densities. In this paper, we evaluate a computationally feasible alternate form, which we call a bi-level MHT algorithm. The basic form of this algorithm has been previously proposed, but results on its performance have been lacking. In addition to describing an implementation of a bi-level MHT algorithm, this paper present Monte Carlo simulation results characterizing the performance of the algorithm, and demonstrates the tradeoff between track acquisition range and false track rate for a simple IRST fly-by scenario.

A widely used estimator for multi-target tracking, conventionally referred to as a maximum likelihood estimator, is analyzed. For the problem of locating closely spaced fixed, untagged objects using a noiseless sensor, the conventional estimator's mean and variance estimates are inconsistent, i.e. asymptotically biased. We propose an alternative maximum likelihood estimator that corrects this problem. This estimator uses a coherent sum over report to track associations to evaluate the track likelihood function. The resulting estimator is efficient in the sense that it achieves the Cramer-Rao lower bound (CRLB) on the variance asymptotically. A novel feature of this approach is that it entails estimation of track error correlations in addition to the variance estimates generated in the usual Kalman filter based methods. These motions are used to develop a filter for a pair of uncorrelated Brownian walkers. It successfully estimates the error correlations that must be present in an optimal filter.

Probabilistic data association approaches are described for tracking multiple targets. These approaches employ multiple frames of data in the data association processing. These approaches offer improved performance over Joint Probabilistic Data Association tracking. This improved performance is obtained, however, at the expense of increased processing load. In the algorithms are design parameters that can be selected to adjust to suit a specific application. The concept of retrodicted hypothesis probability is introduced. Retrodicted hypothesis probabilities are used in an effort to better approximate optimal tracking. Some of these algorithms are retrospective in that, as each new frame of sensor data becomes available, earlier tracks are modified and these changes impact subsequent tracks.

This is a description of a model-based maximum likelihood estimation technique for determining the position and intensities of closely spaced objects (CSO's) present in the focal plane of a forward-looking infrared (FLIR) sensor. The object model considered here is approximate point sources; we present a methodology to superresolve two point sources separated closer than the Rayleigh resolution criteria. The Cramer-Rao theoretical lower bound is derived in closed form; the variance of the proposed estimator will be compared to this bound to verify its superresolving capability. Simulation results are presented for medium and high signal-to-noise (Gaussian noise) ratios and source separations.

In this paper we address the problem of tracking a signal through successive intervals of matched-filter processing. The common approach to this problem is to locate candidate detections in the matched-filter output at each interval, to associate successive detections in state space, to estimate successive states through a Kalman filter application, and to rank association sequences (tracks) with respect to kinematic consistency. However, if a signature model is available and matched-filter statistics are known, the matched-filter output can be converted to a likelihood function that can drive recursive Bayesian processing for the signal state distribution (and no-signal probability). The field tracker described here follows this processing, compromising the optimality of the Bayesian approach only through the discreteness of the state-space domain. The field approach has a detection capability superior to that of an association approach for low SNR signals, and it is highly compatible with parallel processing. An application to the detection and tracking of a constant-velocity signal in 4-D state space (2-D position, 2-D velocity) is provided by way of illustration, and it demonstrates the ability to achieve a conclusive detection of a 3.5-dB-per-update target in ten updates. A number of application alternatives are described that extend the concept to multiple-target scenarios, refined velocity estimates, connection to velocity-independent processing steams, and computationally efficient means of estimating kinematic variables and signal amplitude through auxiliary fields.

Tracking under challenging conditions is a relative new field. Challenging conditions include target maneuvers, close multiple targets or excessive false signals. As in any rapidly growing field, many terms are used in this field that have a specialized meaning and the definitions are not universal. Recognizing the need to improve communications in this field, the SDI Panels on Tracking undertook the task of developing a glossary of terms related to target tracking. The emphasis in this glossary is on terms and definitions unique to single and multiple target tracking.