This PDF file contains the front matter associated with SPIE Proceedings Volume 6969, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and the Conference Committee listing.

Effective missile warning and countermeasures remain an unfulfilled goal for the Air Force and others in the DOD community. To make the expectations a reality, newer sensors exhibiting the required sensitivity, field of regard, and spatial resolution are being developed and transitioned. The largest concern is in the first stage of a missile warning system: detection, in which all targets need to be detected with a high confidence and with very few false alarms. Typical fielded sensors are limited in their detection capability by either lack of sensitivity or by the presence of heavy background clutter, sun glints, and inherent sensor noise. Many threat environments include false alarm sources like burning fuels, flares, exploding ordinance, arc welders, and industrial emitters. Multicolor discrimination has been shown as one of the effective ways to improve the performance of missile warning sensors, particularly for heavy clutter situations. Its utility has been demonstrated in multiple demonstration and fielded systems. New exploitations of background and clutter spectral contents, coupled with advanced spatial and temporal filtering techniques, have resulted in a need to have a new baseline algorithm on which future processing advances may be judged against. This paper describes the AFRL Suite IIIc algorithm chain and its performance against long-range dim targets in clutter.

A Chem/Bio Defense Algorithm Benchmark is proposed as a way to leverage algorithm expertise and apply it to high fidelity Chem/Bio challenge problems in a high fidelity simulation environment. Initially intended to provide risk mitigation to the DTRA-sponsored US Army CUGR ACTD, its intent is to enable the assessment and transition of algorithms to support P3I of future spiral updates. The key chemical sensor in the CUGR ACTD is the Joint Contaminated Surface Detector (JCSD), a short-range stand-off Raman spectroscopy sensor for tactical in-the-field applications. The significant challenges in discriminating chemical signatures in such a system include, but are not limited to, complex background clutter and low signal to noise ratios (SNR). This paper will present an overview of the Chem-Bio Defense Algorithm Benchmark, and the JCSD Challenge Problem specifically.

Because of the unique Raman spectrum of a chemical, Raman spectroscopy can be used to identify chemicals on a surface. In this paper chemical detection and classification in a stationary background are addressed. Firstly, because the autoregressive (AR) spectrum is capable of representing a wide range of spectra, both the pure background and background plus a chemical are modeled as AR spectra with different coefficients. Based on this modeling, a generalized likelihood ratio test (GLRT) is proposed to detect abnormal chemicals in the background. In essence, the GLRT detector tests if the data can be represented by a known AR background spectrum. With the AR spectrum modeling, a classifier based on the locally most powerful test is also proposed to classify the detected chemicals. Computer simulation results are given, which show the effectiveness of the proposed algorithms. Practical problems, such as setting the detection threshold, extension to nonstationary backgrounds, and the identifiability of chemicals are also discussed.

This paper uses super-resolution methods to detect small objects in infrared image sequences from a simulated airborne platform, using image registration techniques for automatic sightline stabilisation. The scene consists of multiple layers, corresponding to a static background scene and layers of cloud cover at varying heights. The motivation is to evaluate the performance of super-resolution methods in the presence of three-dimensional structured infrared clutter.

Distributed Sensor Concept - DISCO was proposed for multiplication of individual sensor capabilities through non-coherent cooperative target engagement. The signal processing technique for DISCO is Recursive Adaptive Frame Integration of Limited data - RAFIL technique that was initially proposed as a way to improve the SNR, reduce data rate and mitigate FPA noise for IR sensors. In DISCO, the RAFIL technique is used in a segmented way, when constituencies of the technique are spatially and temporally separated between individual sensors. Each sensor provides to and receives data from other sensors in the network. In this paper efficiency of DISCO is discussed for acquisition, accurate handover and track correlation of small targets.

Hyperspectral images provide scientists and engineers with the capability of precise material identification in remote sensing applications. One can leverage this data for precise track identification (ID) and incorporate the high-confidence ID in the tracking process. Our previous work demonstrates that hyperspectral-aided tracking outperforms kinematic-only tracking where multiple ambiguous situations exist. We develop a novel gating concept for hyperspectral measurements, similar in concept to the gating of the Mahalanobis distance computed from the Kalman residuals. Our spectral gating definition is based on the distance between the spectral distribution of the class ID of a track and the spectral distribution of the class ID resulting from the classification of a measurement. We further incorporate the distance between each class distribution (in spectral space) in the track association portion of our hyperspectral-aided tracker. Since functional forms of the joint probability distribution function do not exist, similarity measures such as the Kullback-Leibler divergence or Bhattacharyya distance cannot be used. Instead, we compute all pair-wise distances between all samples of the two classes and then summarize these distances in a meaningful way. This article presents our novel spectral gating approach and its use in track association. It further explores different similarity measures and their effect on spectral gating and track association.

Differentiation between particulate biological agents and non-biological agents is typically performed via a time-consuming "wet chemistry" process or through the use of fluorescent and spectroscopic analysis. However, while these methods can provide definitive recognition of biological agents, many of them have to be performed in a laboratory environment, or are difficult to implement in the field. Optical recognition techniques offer an additional recognition approach that can provide rapid analysis of a material in-situ to identify those materials that may be biological in nature. One possible application is to use these techniques to "screen" suspicious materials and to identify those that are potentially biological in nature. Suspicious materials identified by this screening process can then be analyzed in greater detail using the other, more definitive (but time consuming) analysis techniques. This presentation will describe the results of a feasibility study to determine whether optical pattern recognition techniques can be used to differentiate biological related materials from non-biological materials. As part of this study, feature extraction algorithms were developed utilizing multiple contrast and texture based features to characterize the macroscopic properties of different materials. In addition, several pattern recognition approaches using these features were tested including cluster analysis and neural networks. Test materials included biological agent simulants, biological agent related materials, and non-biological materials (suspicious white powders). Results of a series of feasibility tests will be presented along with a discussion of the potential field applications for these techniques.

This paper describes a novel set of algorithms that allows indoor activity to be monitored using data from very low resolution imagers and other non-intrusive sensors. The objects are not resolved but activity may still be determined. This allows the use of such technology in sensitive environments where privacy must be maintained. Spectral un-mixing algorithms from remote sensing were adapted for this environment. These algorithms allow the fractional contributions from different colours within each pixel to be estimated and this is used to assist in the detection and monitoring of small objects or sub-pixel motion.

Operating in a coastal environment, with a multitude of boats of different sizes, detection of small extended targets is only one problem. A further difficulty is in discriminating detections of possible threats from alarms due to sea and coastal clutter, and from boats that are neutral for a specific operational task. Adding target features to detections allows filtering out clutter before tracking. Features can also be used to add labels resulting from a classification step. Both will help tracking by facilitating association. Labeling and information from features can be an aid to an operator, or can reduce the number of false alarms for more automatic systems. In this paper we present work on clutter reduction and classification of small extended targets from infrared and visual light imagery. Several methods for discriminating between classes of objects were examined, with an emphasis on less complex techniques, such as rules and decision trees. Similar techniques can be used to discriminate between targets and clutter, and between different classes of boats. Different features are examined that possibly allow discrimination between several classes. Data recordings are used, in infrared and visual light, with a range of targets including rhibs, cabin boats and jet-skis.

We present a method for detecting a large number of moving targets, such as cars and people, in geographically referenced video. The problem is difficult, due to the large and variable number of targets which enter and leave the field of view, and due to imperfect geo-projection and registration. In our method, we assume feature extraction produces a collection of candidate locations (points in 2D space) for each frame. Some of these locations are real objects, but many are false alarms. Typical feature extraction might be frame differencing, or target recognition. For each candidate location, and at each time step, our algorithm outputs a velocity estimate and confidence which can be thresholded to detect objects with constant velocity. In this paper we derive the algorithm, investigate the free parameters, and compare its performance to a multi-target tracking algorithm.

The various asymmetrical threats in the urban environment have driven the need for persistent surveillance and methods to exploit the data provided by passive sensing platforms. The primary goal is to track vehicles as they move through the urban environment. The rather large number of ambiguous tracking events requires incorporation of target features to maintain track purity. This paper will discuss a feature extraction technique that will be referred to as "feature-aided" tracking to mitigate some of the tracking issues in this environment (e.g. rotation and illumination invariance, partial occlusion, and move-stop-move transitions). The feature extraction method applied is loosely based on the SPIN histogram method of applying a two-dimensional histogram relative to the center of an object. This paper focuses on applying a simplified version of the intensity-based two-dimensional histogram and gradient-based two-dimensional histogram introduced by the works of Mikolajczyk and Schmid, and Lazebnik, Schmid, and Ponce. Instead of applying the matching technique on a still frame subjected to various image transformations, we will apply this technique to sequential frames of imagery in an urban environment. This approach is intended to be the first of several steps towards eventually integrating a feature-aided tracking option as one of multiple sources of measurement association. The preliminary results show potential signs of success especially with rotation-invariance and move-stop-move transitions; however, additional efforts are required associated with illumination invariance, partial occlusion and disambiguation of close proximity objects.

In this paper, a new condition for the target is proposed to increase the robustness of the facet-based detection method for zero-mean Gaussian noise. In the proposed algorithm, the pixels detected from the maximum extremum condition are checked further to discern if they are false maximum points in the proposed scheme. The experimental results show that the proposed algorithm is much more robust for zero-mean Gaussian noise than the conventional detection method.

Bias introduced due to noisy point estimates being propagated through deterministic nonlinear mappings is a reoccurring problem in high-fidelity tracking and classification systems. This paper proves that it is a misconception that such bias is reduced when computing the expected value of the nonlinear output that follows when treating the input as a random vector with expectation equal to the provided estimate. Instead, this doubles the bias. An approximately unbiased estimator and an estimate of its covariance matrix are provided. The estimator can be calculated also in the case where the Hessian matrices associated with the nonlinear mapping are unavailable.

In this paper, we present an algorithm for determining a velocity probability distribution prior from low frame rate aerial video of an urban area, and show how this may be used to aid in the multiple target tracking problem, as well as to provide a foundation for the automated classification of urban transportation infrastructure. The algorithm used to develop the prior is based on using a generic interest point detector to find automobile candidate locations, followed by a series of filters based on scale and motion to reduce the number of false alarms. The remaining locations are then associated between frame pairs using a simple matching algorithm, and the corresponding tracks are then used to build up velocity histograms in the areas that are moved through between the track endpoints. The algorithm is tested on a dataset taken over urban Tucson, AZ. The results demonstrate that the velocity probability distribution prior can be used to infer a variety of information about road lane directions, speed limits, etc..., as well as providing a means of describing environmental knowledge about traffic rules that can be used in tracking.

In order to accurately identify ground targets from radar observations on distributed airborne sensors, range and range-rate measurement data must be either processed onboard the aircraft or at a common control station. This paper will show analysis and results that examine the ability of multiple sensors to provide observability of moving targets. Extremely accurate states are required to support imaging algorithms used to discriminate military targets from civilian targets. Accurate imaging of targets of interest requires sub-meter range accuracy as well as precise knowledge of the target heading which is related to the velocity vector accuracy. The tracking algorithms must provide range accuracy on the order of meters depending on the target spacing and scenario; the imaging pre-processing algorithms can reduce this error to sub-meter levels. Stringent requirements on heading accuracy may be obviated by the use of prominent point tracking.

An assurance region at level p, A_P=p, is an area in motion space that contains the target with assigned probability p. It is on the basis of A_P=p that an action is taken or a decision made. Common model-based trackers generate a synthetic distribution function for the kinematic state of the target. Unfortunately, this distribution is very coarse, and the resulting A_P=p lack credibility. It is shown that a map-enhanced, multiple model algorithm reduces the tracking error and leads to a compact assurance region.

This paper presents a novel continuous approximation approach to nonlinear/non-Gaussian Bayesian tracking. A good representation of the probability density and likelihood functions is essential for the effectiveness of nonlinear filtering algorithms since these functions could be multi-modal. The proposed approach uses B-spline interpolation to represent the density and likelihood functions and tensor product approaches to extend the filter to multidimensional case. The filter is applicable under most general circumstances since it does not make any assumption on the form of the underlying probability density. An advantage of the proposed method is that it retains accurate density information in a continuous low-order polynomial form and finding the target probability in any region of the state space is straightforward. Further processing based on probability density such as finding the higher order moments of the state estimates could also be performed with less computational power. Simulation results are presented to demonstrate the proposed algorithm.

Video cameras onboard multiple unmanned aerial vehicles (UAVs) can provide effective and inexpensive tracking and surveillance functions for ground targets. In our previous work, we quantified the degree of nonlinearity (DoN) of the video filtering problem by considering the perspective transformation for the video measurement model and constant velocity motion for the target dynamic model. In this paper, we generalize the formulation by using a more realistic video measurement model which is based on the perspective transformation, radial and tangential lens distortions, scale, offset, and skew. The centroid pixel coordinates of a target in the digital image represent the sensor measurement for this model. This measurement model is commonly used in photogrammetry, computer vision, and video tracking, where significant height variation can occur. Since the measurement model is a nonlinear function of the target state, the filtering problem is nonlinear. We quantify the DoN of the video filtering problem by calculating the differential geometry based parameter-effects curvature and intrinsic curvature. These measures help a filter designer to select an appropriate nonlinear filtering algorithm for the video filtering problem so that tracking accuracy and computational load requirements are satisfied. Our results show that the DoN of the video filtering problem is quite low and hence a computationally simple filter such as the extended Kalman filter (EKF) is a better choice than the particle filter (PF) which has a much higher computational cost. The state estimation accuracies of the EKF and PF are nearly the same.

In this paper, both Pareto game theory and learning theory algorithms are utilized in a resource management module for a practical missile interception system. The resource management module will determine how many and which antimissiles will be launched for interception. Such interception decisions are based on the number of invading missiles, availability of antimissiles, special capability of antimissiles, and realistic constraints on the movements of both invading missiles and antimissiles such as minimum turning radius, maximum velocity, fuel range, etc. Simulations demonstrate performance improvements when compared to existing strategies (i.e. random assignment), independent of guidance laws (i.e. Proportional Navigation (PN) or the Differential-Game-based Guidance Law (DGL) guidance laws) under end-game interception cases or midcourse interception situations.

In this paper, we consider the tracking of multiple targets in the presence of clutter with poorly localized sensors in multistatic sensor networks. In multistatic sensor networks, we have a few active sensors that emit the signals and many passive sensors that receive the signals originated from the active sensors and reflected by the targets and clutter. In anti-submarine warfare, sensors are typically deployed from aircraft. Optimal tracking performance can be achieved if all the sensor locations are known. However, in general, sensor deployment accuracy is poor, and sensors can also drift significantly over time. Hence, the location uncertainties will increase with time. If the sensors have global position system (GPS) receiver, then their locations can be located with reasonable accuracy. However, most of the cheap sensors do not have a GPS, and therefor, location uncertainties must be taken in to consideration while tracking. An advantage of multistatic sensors compared to independent monostatic sensors is that the sensors can also be tracked accurately. In this paper, we propose how to improve the tracking performance of multiple targets by incorporating sensor uncertainties. We obtain a bound on the tracking performance with location uncertainties being taken into consideration, and propose a technique to select a subset of sensors (if only a few of the available sensors can be used at any measurement time) that should be used at each time step based on the bound. Simulation results illustrating the performance of the proposed algorithms are also presented.

A passive coherent location (PCL) system exploits the ambient FM radio or television signals from powerful local transmitters, which makes it ideal for covert tracking. In a passive radar system, also known as PCL system, a variety of measurements can be used to estimate target states such as direction of arrival (DOA), time difference of arrival (TDOA) or Doppler shift. Noise and the precision of DOA estimation are main issues in a PCL system and methods such as conventional beam forming (CBF) algorithm, algebraic constant modulus algorithm (ACMA) are widely analyzed in literature to address them. In practical systems, although it is necessary to reduce the directional ambiguities, the placement of receivers closed to each other results in larger bias in the estimation of DOA of signals, especially when the targets move off bore-sight. This phenomenon leads to degradation in the performance of the tracking algorithm. In this paper, we present a method for removing the bias in DOA to alleviate the aforementioned problem. The simulation results are presented to show the effectiveness of the proposed algorithm with an example of tracking airborne targets.

Conventional algorithms for track association (termed "correlation" by convention) employ algorithms which are applied to all sensor tracks at a specific time. The overall value of sensor networks for data fusion is closely tied to the reliability of correct association of common objects tracked by the sensors. Multisensor architectures consisting of gaps in target coverage requires that tracks must be propagated substantially forward or backward to a common time for correlation. This naturally gives rise to the question: at which time should track correlation be performed? In the conventional approach, a two-sensor correlation problem would be solved by propagating the first sensor's tracks forward to the update time (current time) of the tracks from the second sensor. We question this approach by showing simulation results that indicate that the current time can be the worst time to correlate. In addition, a methodology for calculating the approximate optimal correlation time for linear-Gaussian tracking problems is provided.

A framework of simultaneously estimating the motion and structure parameters of a 3D object by using high range resolution (HRR) and ground moving target indicator (GMTI) measurements with template information is given. By decoupling the motion and structure information and employing rigid-body constraints, we have developed the kinematic and measurement equations of the problem. Since the kinematic system is unobservable by using only one scan HRR and GMTI measurements, we designed an architecture to run the motion and structure filters in parallel by using multi-scan measurements. Moreover, to improve the estimation accuracy in large noise and/or false alarm environments, an interacting multi-template joint tracking (IMTJT) algorithm is proposed. Simulation results have shown that the averaged root mean square errors for both motion and structure state vectors have been significantly reduced by using the template information.

Due to the availability of cheap passive sensors, it is possible to deploy a large number of them for tracking purposes in anti-submarine warfare (ASW). However, modern submarines are quiet and difficult to track with passive sensors alone. Multistatic sensor networks, which have few transmitters (e.g., dipping sonars) in addition to passive receivers, have the potential to improve the tracking performance. We can improve the performance further by moving the transmitters according to existing target states and any possible new targets. Even though a large number of passive sensors are available, due to frequency, processing power and other physical limitations, only a few of them can be used at any one time. Then the problems are to decide the path of the transmitters and select a subset from the available passive sensors in order to optimize tracking performance. In this paper, the PCRLB, which gives a lower bound on estimation uncertainty, is used as the performance measure. We present an algorithm to decide jointly the optimal path of the movable transmitters, by considering their operational constraints, and the optimal subset of passive sensors that should be used at each time steps for tracking multiple, possibly time-varying, number of targets. Finding the optimal solution in real time is difficult for large scale problems, and we propose a genetic algorithm based suboptimal solution technique. Simulation results illustrating the performance of the proposed algorithm are also presented.

We consider target detection and tracking of stealthy targets. These targets can be characterized by a strong aspect dependence leading to difficult detectability without a multi-static setup. Even in a multi-static setup only sensors in a certain zone can detect the return signal, if the the aspect dependent return has a small bandwidth. We propose a solution based on a large number of simple sensor, as using many receivers increases the probability of detection. The sensors are simple in the sense that they only transmit binary detection results to a fusion center that has comparatively deep capabilities, and they do not need to know their own position or communicate with other sensors. We characterize the target position estimation performance using the Cramer-Rao bound and simulation results, considering uncertainty in nuisance parameters as the sensor positions or the specifics of the aspect dependence. We suggest a data collection protocol that includes locating sensors that detect the target and has low communication complexity. As a novelty we also include information about "non-localized" sensors, as sensors which do not detect the target stay quiet to save bandwidth and energy, therefore are not known to the fusion center except via knowledge of the deployed sensor density and deployment region.

The central problem in multitarget, multisensor surveillance is that of determining which reports from separate sensors arise from common objects. Due to stochastic errors in the source reports, there may be multiple data association hypotheses with similar likelihoods. Moreover, established methods for performing data association make fundamental modeling assumptions that hold only approximately in practice. For these reasons, it is beneficial to include some measure of uncertainty, or ambiguity, when reporting association decisions. In this paper, we perform an analysis of the benefits versus runtime performance of three methods of producing ambiguity estimates for data association: enumeration of the k-best data association hypotheses, importance sampling, and Markov Chain Monte Carlo estimation. In addition, we briefly examine the sensitivity of ambiguity estimates to violations of the stochastic model used in the data association procedure.

When compared to tracking airborne targets, tracking ground targets on urban terrains brings a new set of challenges. Target mobility is constrained by road networks, and the quality of measurements is affected by dense clutter, multipath, and limited line-of-sight. We investigate the integration of detection, signal processing, tracking, and scheduling by exploiting distinct levels of diversity: (1) spatial diversity through the use of coordinated multistatic radars; (2) waveform diversity by adaptively scheduling the transmitted radar waveform according to the scene conditions; and (3) motion model diversity by using a bank of parallel filters, each one matched to a different maneuvering model. Specifically, at each scan, the waveform that yields the minimum one-step-ahead error covariance matrix determinant is transmitted; the received signal is then matched-filtered, and quadratic curve fitting is applied to extract range and azimuth measurements that are input to the LMIPDA-VSIMM algorithm for data association and filtering. Monte Carlo simulations are used to demonstrate the effectiveness of the proposed system on a realistic urban scenario. A more traditional open-loop system, in which waveforms are scheduled on a round-robin fashion and with no other modes of diversity available, is used as a baseline for comparison. Simulation results show that our closed-loop system significantly outperforms the baseline system, presenting both a reduction on the number of lost tracks, and a reduction on the volume of the estimation uncertainty ellipse. The interdisciplinary nature of this work highlights the challenges involved in designing a closed-loop active sensing platform for next-generation urban tracking systems.

Surveillance and ground target tracking using multiple electro-optical and infrared video sensors onboard unmanned aerial vehicles (UAVs) has drawn a great deal of interest in recent years. We compare a number of track-to-track fusion algorithms using a single target with the nearly constant velocity dynamic model and two UAVs. A local tracker is associated with each UAV and processes video measurements to produce local tracks. The video measurement is the centroid pixel location in the digital image corresponding to the target positions on the ground. In order to handle arbitrary height variations, we use the perspective transformation for the video measurement model. In addition, the video measurement model also includes radial and tangential lens distortions, scale, and offset. Since the video measurement model is a nonlinear function of the target position, the tracking filter uses a nonlinear filtering algorithm. A fusion center fuses track data received from two local trackers. The track-to-track fusion algorithms employed by the fusion center include the simple convex combination fusion, Bhattacharya fusion, Bar-Shalom-Campo fusion, and extended information filter based fusion algorithms. We compare the fusion accuracy, covariance consistency, bias in the fused estimate, communication load requirements, and scalability. Numerical results are presented using simulated data.

With current processing power, Multiple Hypothesis Tracking (MHT) becomes a feasible and powerful solution; however a good hypothesis pruning method is mandatory for efficient implementation. The availability of a continuously increasing number of tracking systems raises interest in combining information from these systems. The purpose of this paper is to propose a method of information fusion for such trackers that use MHT locally with local information sent in the form of sensor global hypotheses and the fusion center combining them into fused global hypotheses. The information extracted from the best fused global hypotheses, in the form of ranking of received sensor global hypotheses, is sent back to local trackers, for optimized pruning. Details of the method, in terms of sensor global hypotheses generation, evaluation, pruning at local sensors, association and fusion of sensor global hypotheses at fusion center, and usage of the information received as feedback from the fusion center are presented.

Multi-sensor tracking holds the potential for improving the surveillance performance achieved through single-sensor tracking. This potential has been demonstrated in many domains: at NURC, in the context of multi-static undersea surveillance. Nonetheless, the issue remains of how best to process data in large sensor networks. This issue is taken up in this paper. We are interested to compare multi-sensor scan-based tracking with a two-stage approach: static fusion followed by scan-based tracking. This paper focuses on some candidate methodologies for static fusion. The methods developed in this paper fall into two categories. The scan-based approach leverages the Gaussian mixture probabilistic hypothesis density (GM-PHD) filter; the batch approaches are based on scan statistics, and on the multi-hypothesis PDA (MHPDA). Preliminary simulation-based performance analysis suggests that the MHPDA approach to static fusion is the most robust in dealing with closely spaced targets and small sensor networks. Leveraging the results presented here, follow-on work will address the determination of an optimal fusion and tracking architecture. In particular, we will test scan-based tracking based on the NURC distributed multi-hypothesis tracker (DMHT), with MHPDA processing followed by scan-based tracking (with the DMHT). We anticipate that, for large sensor networks, the latter approach will outperform the former.

Track-to-track fusion (T2TF) is very important in distributed tracking systems. When tracks of a target at different sensors are fused for increased accuracy, an important issue is to account for the crosscorrelations among the tracks. In this paper, an exact solution for the general problem of T2TF is proposed. It can be used with various information structures, e.g., memoryless T2TF or sequential T2TF with information feedback at arbitrary times. Simulation results for a 1-D tracking scenario evaluate the benefit of the various configurations for T2TF. It is also observed that T2TF, although done optimally, can be suboptimal w.r.t. centralized measurement fusion. This is because the locally optimal filter gains are, in general, globally suboptimal. Furthermore, it is shown that feedback can lead to degradation of the accuracy of the (optimally) fused tracks. Based on the exact T2TF algorithms, an approximate implementation which requires less communications between the fusion center and the local trackers is also proposed. This allows the algorithms to be implemented in distributed tracking systems with low communication capacity. Examples of tracking in two dimensions with two radars, show that the proposed T2TF algorithms are consistent and can provide significant improvement in accuracy over unfused tracks. For the sensors-target geometry considered, the T2TF algorithm can even meet the performance bound of the centralized measurement fusion at the fusion times.

The primary components of a target track are the estimated state vector and its error variance-covariance matrix (or simply the covariance). The estimated state indicates the location and motion of the target. The track covariance is intended to indicate the uncertainty or inaccuracy of the target state estimate. The covariance is computed by the track processor and may or may not realistically indicate the inaccuracy of the state estimate. Covariance Consistency is the property that a computed variance-covariance matrix realistically represents the covariance of the actual errors of the estimate. The computed covariance of the state estimation error is used in the computations of the data association processing function and the estimation filter; consequently, degraded track consistency might cause misassociations (correlation errors) and degraded filter processing that can degrade track performance. The computed covariance of the state estimation error is also used by downstream functions, such as the network-level resource management functions, to indicate the accuracy of the target state estimate. Hence, degraded track consistency can mislead those functions and the war fighter about accuracy of each target track. In the development of target trackers, far more attention has been given to improving the accuracy of the estimated target state than in improving the track covariance consistency. This paper addresses covariance compensation to reduce the degradation of consistence due to potential misassociations in measurement fusion using single-frame data association. This compensation approach used is also applicable to other fusion approaches and to tracking with data from a single sensor. This paper also shows how this compensation approach can be applied to a variety of data association algorithms.

A common problem in video-based tracking of urban targets is occlusion due to buildings and vehicles. Fortunately, when multiple video sensors are present with enough geometric diversity, track breaks due to temporary occlusion can be substantially reduced by correlating and fusing source-level track data into system-level tracks. Furthermore, when operating in a communication-constrained environment, it is preferable to transmit track data rather than either raw video data or detection measurements. To avoid statistical correlation due to common prior information, tracklets can be formed from the source tracks prior to transmission to a central command node, which is then responsible for system track maintenance via correlation and fusion. To maximize the operational benefit of the system-level track picture, it should be distributed in an efficient manner to all platforms, especially the local trackers at the sensors. In this paper, we describe a centralized architecture for multi-sensor video tracking that uses tracklet-based feedback to maintain an accurate and complete track picture at all platforms. We will also use challenging synthetic video data to demonstrate that our architecture improves track completeness, enhances track continuity (in the presence of occlusions), and reduces track initiation time at the local trackers.

The joint target tracking and classification using target-to-sensor aspect-dependent Radar Cross Section (RCS) and kinematic data for multistatic sonar network is presented in this paper. The scattered signals measured from different orientations of a target may vary due to aspect-dependant RCS. A complex target may contain several dozen significant scattering centers and dozens of other less significant scatterers. Because of this multiplicity of scatterers, the net RCS pattern exhibits high variation with aspect angle. Thus, radar cross sections from multiple aspects of a target, which are obtained via multiple sensors, will help in accurately determining the target class. By modeling the deterministic relationship that exits between RCS and target aspect, both the target class information and the target orientation can be estimated. Kinematic data are also very helpful in determining the target class as it describes the target motion pattern and its orientation. The proposed algorithm exploits the inter-dependency of target state and the target class using aspect-dependent RCS and kinematic information in order to improve both the state estimates and classification of each target. The simulation studies demonstrate the merits of the proposed joint target tracking and classification algorithm based on aspect-dependant RCS and kinematic information.

This research investigates the impact of scene context knowledge on tracking vehicles in an urban environment based on video image change detection. The scene context consists of knowledge of the road network and 3D building properties. Airborne sensor position information relative to a 3D model of the context enables calculation of building occlusions of ground locations. From this context, probability of detection maps that include regions of interest and smoothed lines-of-sight are developed that assist the change detection algorithm in reducing false alarms.

In distributed tracking systems, multiple non-collocated trackers cooperate to fuse local sensor data into a global track picture. Generating this global track picture at a central location is fairly straightforward, but the single point of failure and excessive bandwidth requirements introduced by centralized processing motivate the development of decentralized methods. In many decentralized tracking systems, trackers communicate with their peers via a lossy, bandwidth-limited network in which dropped, delayed, and out of order packets are typical. Oftentimes the decentralized tracking problem is viewed as a local tracking problem with a networking twist; we believe this view can underestimate the network complexities to be overcome. Indeed, a subsequent 'oversight' layer is often introduced to detect and handle track inconsistencies arising from a lack of robustness to network conditions. We instead pose the decentralized tracking problem as a distributed database problem, enabling us to draw inspiration from the vast extant literature on distributed databases. Using the two-phase commit algorithm, a well known technique for resolving transactions across a lossy network, we describe several ways in which one may build a distributed multiple hypothesis tracking system from the ground up to be robust to typical network intricacies. We pay particular attention to the dissimilar challenges presented by network track initiation vs. maintenance and suggest a hybrid system that balances speed and robustness by utilizing two-phase commit for only track initiation transactions. Finally, we present simulation results contrasting the performance of such a system with that of more traditional decentralized tracking implementations.

The primary components of a target track are the estimated state vector and its error variance-covariance matrix (or simply the covariance). The estimated state indicates the location and motion of the target. The track covariance is intended to indicate the uncertainty or inaccuracy of the target state estimate. The covariance is computed by the track processor and may or may not realistically indicate the inaccuracy of the state estimate. Covariance Consistency is the property that a computed variance-covariance matrix realistically represents the covariance of the actual errors of the estimate. The computed covariance of the state estimation error is used in the computations of the data association processing function and the estimation filter; consequently, degraded track consistency might cause misassociations (correlation errors) and degraded filter processing that can degrade track performance. The computed covariance of the state estimation error is also used by downstream functions, such as the network-level resource management functions, to indicate the accuracy of the target state estimate. Hence, degraded track consistency can mislead those functions and the war fighter about accuracy of each target track. In the development of target trackers, far more attention has been given to improving the accuracy of the estimated target state than in improving the track covariance consistency. This paper addresses covariance compensation to reduce the degradation of consistence due to potential misassociations in measurement fusion using single-frame data association. This compensation approach used is also applicable to other fusion approaches and to tracking with data from a single sensor. This paper presents simplifications in some of the processing of the covariance compensation to reduce the processing complexity, i.e. processor load.

In this paper, we present an efficient data association algorithm for tracking ground targets that perform move-stop-move maneuvers using ground moving target indicator (GMTI) radar. A GMTI radar does not detect the targets whose radial velocity falls below a certain minimum detectable velocity. Hence, to avoid detection enemy targets deliberately stop for some time before moving again. When targets perform move-stop-move maneuvers, a missed detection of a target by the radar leads to an ambiguity as to whether it is because the target has stopped or due to the probability of detection being less than one. A solution to track move-stop-move target tracking is based on the variable structure interacting multiple model (VS-IMM) estimator in an ideal scenario (single target tracking with no false measurements) has been proposed. This solution did not consider the data association problem. Another solution, called two-dummy solution, considered the data association explicitly and proposed a solution based on the multiframe assignment algorithm. This solution is computationally expensive, especially when the scenario is complex (e.g., high target density) or when one wants to perform high dimensional assignment. In this paper, we propose an efficient multiframe assignment-based solution that considers the second dummy measurement as a real measurement than a dummy. The proposed algorithm builds a less complex assignment hypothesis tree, and, as a result, is more efficient in terms of computational resource requirement.

We describe a new nonlinear filter that is vastly superior to the classic particle filter. In particular, the computational complexity of the new filter is many orders of magnitude less than the classic particle filter with optimal estimation accuracy for problems with dimension greater than 2 or 3. We consider nonlinear estimation problems with dimensions varying from 1 to 20 that are smooth and fully coupled (i.e. dense not sparse). The new filter implements Bayes' rule using particle flow rather than with a pointwise multiplication of two functions; this avoids one of the fundamental and well known problems in particle filters, namely "particle collapse" as a result of Bayes' rule. We use a log-homotopy to derive the ODE that describes particle flow. This paper was written for normal engineers, who do not have homotopy for breakfast.

Timely recognition of threats can be significantly supported by security assistance systems that work continuously in time and call the attention of the security personnel in case of anomalies. We describe the concept and the realization of an indoor security assistance system for real-time decision support. Data for the classification of persons are provided by chemical sensors detecting hazardous materials. Due to their limited spatio-temporal resolution, a single chemical sensor cannot localize this material and associate it with a person. We compensate this deficiency by fusing the output of multiple, distributed chemical sensors with kinematical data from laser-range-scanners. Both, tracking and fusion of tracks with chemical attributes can be processed within one single framework called Probabilistic Multiple Hypothesis Tracking (PMHT). An extension of PMHT for dealing with classification measurements (PMHT-c) already exists. We show how PMHT-c can be applied to associate chemical attributes to person tracks. This affords the localization of threads and a timely notification of the security personnel.

This paper presents procedures to calculate the probability that the measurement originating from an extraneous target will be (mis)associated with a target of interest for the cases of Nearest Neighbor and Global association. It is shown that these misassociation probabilities depend, under certain assumptions, on a particular - covariance weighted - norm of the difference between the targets' predicted measurements. For the Nearest Neighbor association, the exact solution, obtained for the case of equal innovation covariances, is based on a noncentral chi-square distribution. An approximate solution is also presented for the case of unequal innovation covariances. For the Global case an approximation is presented for the case of "similar" innovation covariances. In the general case of unequal innovation covariances where this approximation fails, an exact method based on the inversion of the characteristic function is presented. The theoretical results, confirmed by Monte Carlo simulations, quantify the benefit of Global vs. Nearest Neighbor association. These results are applied to problems of single sensor as well as centralized fusion architecture multiple sensor tracking.

Small object detection with a low false alarm rate remains a challenge for automated hyperspectral detection algorithms when the background environment is cluttered. In order to approach this problem we are developing a compact hyperspectral sensor that can be fielded from a small unmanned airborne platform. This platform is capable of flying low and slow, facilitating the collection of hyperspectral imagery that has a small ground-sample distance (GSD) and small atmospheric distortion. Using high-resolution hyperspectral imagery we simulate various ranges between the sensor and the objects of interest. This numerical study aids in analysis of the effects of stand-off distance on detection versus false alarm rates when using standard hyperspectral detection algorithms. Preliminary experimental evidence supports our simulation results.

Sensor jitter introduces non-white noise fluctuations in imagery of cluttered scenes. These fluctuations are a major source of interference in the detection of weak time-dependent signals, which may be associated with a subject's appearance, motion, or brightness modulation. Due to the presence of sensor pattern noise and uncertainty in the scene's subpixel spatial structure, standard frame-to-frame registration methods have limited ability to model and remove these fluctuations. A simple temporal whitening approach, applicable to a wide variety of imaging systems, is found to be highly effective for suppressing subpixel jitter effects, leading to dramatic (up to several orders of magnitude) improvement in signal detection ability.

In this paper a real-time cooperative path decision algorithm for UAV surveillance is proposed. The surveillance mission includes multiple objectives: i) Navigate the UAVs safely in a hostile environment; ii) Search for new targets in the surveillance region; iii) Classify the detected targets; iv) Maintain tracks on the detected targets. To handle these competing objectives, a layered decision framework is proposed, in which different objectives are relevant at different decision layers according to their priorities. Compared to previous work, in which multiple objectives are integrated into a single global objective function, this layered decision framework allows detailed specification of the desired performance for each objective and guarantees that an objective with high priority will be first satisfied by eliminating possible compromises from other less important ones. In addition, path decision strategies that are suited to individual objectives can be used at different decision layers. The layered decision framework, along with a multi-step look-ahead path decision strategy based on a Roll-out policy is shown to be able to guide the UAV group effectively for the multi-objective surveillance in a hostile environment.

Passive coherent location (PCL), which uses the commercial signals as illuminators of opportunity, is an emerging technology in air defense systems. The advantages of PCL are low cost, low vulnerability to electronic counter measures, early detection of stealthy targets and low-altitude detection. However, limitations of PCL include lack of control over illuminators, poor bearing accuracy, time-varying sensor parameters and limited observability. In this paper, multiple target tracking using PCL with high bearing error is considered. In this case, the challenge is to handle high nonlinearity due to high measurement error. In this paper, we implement the converted measurement Kalman filter, unscented Kalman filter and particle filter based PHD filter for PCL radar measurements and compare their performances.

Bias estimation using objects with unknown data association requires concurrent estimation of both biases and optimal data association. This report derives maximum a posteriori (MAP) data association likelihood ratios for concurrent bias estimation and data association based on sensor-level track state estimates and their joint error covariance. Our approach is unique for two reasons. First, we include a bias prior that allows estimation of absolute sensor biases, rather than just relative biases. Second, we allow concurrent bias estimation and association for an arbitrary number of sensors. The two-sensor likelihood ratio is derived as a special case of the general M-sensor result.

In multi-sensor tracking systems, observations are often exchanged over a network for processing. Network delays create situations in which measurements arrive out-of-sequence. The out-of-sequence measurement (OOSM) update problem is of particular significance in networked multiple hypothesis tracking (MHT) algorithms. The advantage of MHT is the ability to revoke past measurement assignment decisions as future information becomes available. Accordingly, we not only have to deal with network delays for initial assignment, but must also address delayed assignment revocations. We study the performance of extant algorithms and two algorithm modifications for the purpose of OOSM filtering in MHT architectures.

This paper presents a new small target detection method using scale invariant feature. Detecting small targets whose sizes are varying is very important to automatic target detection in infrared search and track (IRST). The conventional spatial filtering methods with fixed sized kernel show limited target detection performance for incoming targets. The scale invariant target detection can be defined as searching for maxima in the 3D (x, y, and scale) representation of an image with the Laplacian function. The scale invariant feature can detect different sizes of targets robustly. Experimental results with real FLIR images show higher detection rate and lower false alarm rate than conventional methods. Furthermore, the proposed method shows very low false alarms in scan-based IR images than conventional filters.