This PDF file contains the front matter associated with SPIE Proceedings Volume 6566, including the Title Page, Copyright information, Table of Contents, and the Conference Committee listing.

Exact wave theories of specular reflectance from rough surfaces are computationally intractable thus motivating the practical need for geometric reflectance models which treat only the geometric ray nature of light reflection. The cornerstone of geometric reflectance modeling from rough surfaces in computer vision and computer graphics over the past two decades has been the Torrance-Sparrow model. This model has worked well as an intuitive description of rough surfaces as a collection of planar Fresnel reflectors called microfacets together with the concept of geometric attenuation for light which is obscured during reflection under an assumed rough surface geometry. Experimental data and analysis show that the current conceptualization of how specularly reflected light rays geometrically interact with rough surfaces needs to be seriously revised. The Torrance-Sparrow model while in qualitative agreement with specular reflection from rough surfaces is seen to be quantitatively inaccurate. Furthermore there are conceptual inconsistencies upon which derivation of this reflectance model is based. We show how significant quantitative improvement can be achieved for a geometric reflectance model by making some fundamental revisions to notions of microfacet probability distributions and geometric attenuation. Work is currently undergoing, to relate physical surface reconstructions from Atomic Force Microscope data to reflectance data from these same surfaces.

We examine the sensitivity of minimum noise and correlation energy (MINACE) filters to three different types of distortion variations (aspect view, depression angle, and scale) that are typically present in infrared (IR) imagery used for automatic target recognition (ATR) and tracking applications. Prior DIF (distortion-invariant filter) ATR work has addressed at most two simultaneous variations - aspect view and depression angle variations for SAR data, and aspect view and thermal state variations for IR data. No prior Minace ATR work has addressed scale variations. In our tests, we consider all three simultaneous variations - aspect view, depression angle, and scale. This is new. Our goal is to determine if one Minace filter per object can handle full 360° aspect view variations and can handle small depression angle variations, and to determine the range of scales that one Minace filter per object can handle after training on data at one or more scales. This determines when new Minace filters are needed in an image closing sequence. In all cases, shifts of the target test inputs are considered. We use our autoMinace algorithm that automates selection of the Minace filter parameter c and the training set images to be included in the filter. We also consider rejection of unseen confuser objects and clutter. No confuser, clutter, or test set data are present in the training or the validation set. We present test results using both real and CAD IR data.

Common strategies for detection and tracking of aerial moving targets in airborne Forward-Looking Infrared (FLIR) images offer accurate results in images composed by a non-textured sky. However, when cloud and earth regions appear in the image sequence, those strategies result in an over-detection that increases very signficantly the false alarm rate. Besides, the airborne camera induces a global motion in the image sequence that complicates even more detection and tracking tasks. In this work, an automatic detection and tracking system with an innovative and efficient target trajectory filtering is presented. It robustly compensates the global motion to accurately detect and track potential aerial targets. Their trajectories are analyzed by a curve fitting technique to reliably validate real targets. This strategy allows to filter false targets with stationary or erratic trajectories. The proposed system makes special emphasis in the use of low complexity video analysis techniques to achieve real-time operation. Experimental results using real FLIR sequences show a dramatic reduction of the false alarm rate, while maintaining the detection rate.

To perform multi-sensors simulations, the French DGA/DET (Directorate for Technical Evaluation of the French Ministry of Defense) uses CHORALE (simulated Optronic Acoustic Radar battlefield). CHORALE enables the user to create virtual and realistic multi spectral 3D scenes, and generates the physical signal received by one or several sensors, typically an IR sensor or an acoustic sensor. This article presents how the expertise is made to evaluate smart ammunition to detect ground target with infrared sensor and shape detector in a virtual battlefield with the environment CHORALE and the workshop AMOCO. The scene includes background with trees, houses, roads, fields, targets, and the ammunition. Each tool is explained to understand the physics phenomena in the scene to take into account atmospheric transmission, radiative parameters of objects and counter-measure devices. Then numeric models are described as the 6 DOF ballistics models, sensor model according precise positions inside the ammunition as well as the different steps of calculation between industrial model and technical model to obtain the global simulation. Finally, this paper explains some results of the evaluation compared with the true behavior after tests on proving ground. Then future evolutions are presented to perform similar evaluation with other kind of intelligent ammunition in a real-time model.

Efficient processing of imagery derived from remote sensing systems has become ever more important due to increasing data sizes, rates, and bit depths. This paper proposes a target detection method that uses a special class of wavelets based on highly frequency-selective directional filter banks. The approach helps isolate object features in different directional filter output components. These components lend themselves well to the application of powerful denoising and edge detection procedures in the wavelet domain. Edge information is derived from directional wavelet decompositions to detect targets of known dimension in electro optical imagery. Results of successful detection of objects using the proposed method are presented in the paper. The approach highlights many of the benefits of working with directional wavelet analysis for image denoising and detection.

In this keynote address, we address three-dimensional (3D) distortion-tolerant object recognition using photon-counting integral imaging (II). A photon-counting linear discriminant analysis (LDA) is discussed for classification of photon-limited images. We develop a compact distortion-tolerant recognition system based on the multiple-perspective imaging of II. Experimental and simulation results have shown that a low level of photons is sufficient to classify out-of-plane rotated objects.

Automatic target recognition (ATR) using an infrared (IR) sensor is a particularly appealing combination, because an IR sensor can overcome various types of concealment and works in both day and night conditions. We present a system for ATR on low resolution IR imagery. We describe the system architecture and methods for feature extraction and feature subset selection. We also compare two types of classifier, K-Nearest Neighbors (KNN) and Random Decision Tree (RDT). Our experiments test the recognition accuracy of the classifiers, within our ATR system, on a variety of IR datasets. Results show that RDT and KNN achieve comparable performance across the tested datasets, but that RDT requires significantly less retrieval time on large datasets and in high dimensional feature spaces. Therefore, we conclude that RDT is a promising classifier to enable a robust, real time ATR solution.

The Defense Advanced Research Projects Agency (DARPA) Video Verification of Identity (VIVID) program has as its goal the development of the best video tracker ever. This goal is reached through a philosophy of on-the-fly target modeling and the use of three distinct modules: a multiple-target tracker, a confirmatory identification module, and a collateral damage avoidance/moving target detection module. Over the two years of VIVID Phase I, progress appraisal of the ATR-like confirmatory identification module was provided to DARPA by the Air Force Research Laboratory Comprehensive Performance Assessment of Sensor Exploitation (COMPASE) Center through regular evaluations. This document begins with an overview of the VIVID system and its approach to solving the multiple-target tracking problem. A survey of the data collected under VIVID auspices and their use in the evaluation are then described, along with the operating conditions relevant to confirmatory identification. Finally, the evaluation structure is presented in detail, including metrics, experiment design, experiment construction techniques, and support tools.

An efficient multiple target recognition technique which combines the inherited enhancement of the optical polarization field with the feature enhancement of the wavelet filter is proposed in this paper. The wavelet filter is superimposed on the joint power spectrum before the correlation output is produced. It is shown that the proposed technique yields cumulative target discrimination capabilities for automatic single/multiple target recognition applications.

The ability of interferoceiver to reveal micro Doppler signatures from a single radar pulse opens a new era in automatic target recognition. The present paper considers a real scenario to illustrate the power of such ability. In July 3, 1988, a United States Navy battleship in the Persian Gulf shot down an Iranian passenger plane Flight 655. The Navy said they mistook it for a jet fighter. The incident killed all 290 people on board the passenger plane. This is a case of misidentification, which could be avoided with the help of interferoceiver.

A flying platform illuminates a land mine field with mixtures of various landmines (i.e., buried, on the surface, plastic or metallic) and some "confusers", with an ultra-wideband (UWB) radar. The polarimetric echoes returned by the mine field are mapped into an overall synthetic aperture radar (SAR) image, which is then analyzed pixel-by-pixel by modern time-frequency (t-f) techniques. The t-f analysis of any echo from any of the individual scatterers in the mine field can be performed using a number of t-f distributions, which in turn generate two-dimensional plots of each such scatterer in t-f space. These plots are richer in information than those in the original SAR image, and they offer a larger variety of clues useful for the discrimination of each type of mine from the others or from the confusers. Several t-f distributions are employed in the study, and it is found that some are better than others for the present purpose of target detection and classification. From the images obtained we can conclude that the Pseudo-Wigner-Ville and the Choi-Williams distributions provide the best discrimination results. It is also found that the larger mines such as those denoted here as of "type-1" are the easiest to identify. Using the above-mentioned distributions it follows that the distinction between actual mines and clutter objects (or "confusers") becomes clearer, particularly when the latter objects are metallic. Numerous images generated in this study confirm the above conclusions.

In active sonar or radar, if the channel is spatially- and/or temporally-varying, then the target echo can change with propagation, such that echoes from identical targets may not be identified as such at the receiver. Two common propagation effects that induce changes in the signal are dispersion and dissipation (or damping), which give rise to frequency-dependent velocity of propagation and frequency-dependent attenuation, respectively. We have previously developed a feature extraction process for target echoes in dispersive channels, to obtain moment-like features that are invariant to dispersion, per mode. Accordingly, even though the target echo can change with propagation in a dispersive channel, the "dispersion-invariant moment" features do not. However, these moment features are affected by damping. In this paper, we consider the case of a channel with dispersion and damping, and derive features that are invariant to both phenomena, for any dispersion relation and exponential or power-law damping. Results are presented from classification simulations to demonstrate the utility of these features.

We show how to construct distributions from moments directly, that is, without first calculating the characteristic function. We apply the method to compute the Wigner distribution from its conditional moments.

Pattern recognition in hyperspectral imagery is a challenging task as the objects occupy only a few pixels or less. The presence of noise can make detection more complicated as spectral signature of pixels can change due to noise. In this paper a technique is proposed for detection in hyperspectral imagery using one dimensional maximum average correlation height (MACH) filter. MACH filter is a type of matched spatial training filter which is widely used for spatial aperture radar (SAR), laser radar (LADAR), forward looking infrared (FLIR) and other class of two-dimensional imageries to train and detect objects. For hyperspectral case a modified one-dimensional MACH filter is proposed which uses likely variations of a given ideal spectral signature for training. Each pixel vector of the data cube is then compared with the detection filter using Mahalanobis distance. Based on Mahalanobis distance between the trained filter and the pixels of the imagery, two classes are formed called the background class which does not contain a desired object and the object class which does contain the desired object. By applying threshold boundary, a decision is then made whether a given pixel belongs to the background class or object class. The simulation results using real life hyperspectral imagery show that the proposed technique can detect and classify the desired objects with a higher rate of efficiency even for very small and scattered objects.

In this paper, we proposed a nonlinear unmixing matching algorithm using bidirectional reflectance function (BDRF) and maximum liklihood estimation (MLE). Spectral unmixing algorithms are used to determine the contribution of multiple substances in a single pixel of a hyperspectral image. For any kind of unmixing model basic approach is to describe how different substances are combined in a composite spectrum. When a linear reationship exists between the fractional abundance of the substances, linear unmixing algorithms can determine the endmembers present in that particular pixel. When the relationship is not linear rather each substance is randomly distributed in a homogeneous way the mixing is called nonlinear. Though there are plenty of unmixing algorithms based on linear mixing models (LMM) but very few algorithms have developed to to unmix nonlinear data. We proposed a nonlinear unmixing technique using BDRF and MLE and tested our algorithm using both synthetic and real hyperspectral data.

Now a days detection of man made or natural object using hyperspectral imagery is a great interest of both civilian and military application. With compared to other method, hyperspectral image processing can detect both full pixel and subpixel object by analyzing the fine details of both target and background signatures. There are lots of algorithms to detect hyperspectral full pixel targets. There are also methods to detect subpixel target [1-2]. In this paper we have presented an automated method to detect hyperspectral targets using Linear Mixing Model (LMM) [4]. In our method we estimated the background endmember signatures Vertex Component Analysis which is a fast algorithm to unmix hyperspectral data [6] after removing target like pixels. Sensor noise is modeled as a Gaussian random vector with uncorrelated components of equal variance. This paper provides a complete and self-contained theoretical derivation of a subpixel target detector using the Generalized Likelihood Ratio Test (GLRT) approach and the LMM [4].

Probabilistic graphical models, in particular Bayesian networks, provide a consistent framework in which to address problems containing uncertainty and complexity. Probabilistic inference in high-dimensional problems only becomes tractable when the system can be made modular by imposing meaningful conditional independence assumptions. Bayesian networks provide a natural way to accomplish this. As a combination of probability theory and graph theory, the probabilistic aspects of a graphical model provide a consistent way of connecting data to models, while graph theory provides an intuitively appealing interface to express independence assumptions as well as efficient computation algorithms. A detailed example demonstrating various aspects of Bayesian networks for an electronic intelligence (ELINT) sensor data fusion decision system is presented, including a Value of Information (VOI) analysis.

The problem of target recognition using signatures collected by a multi-radar system is investigated in this paper. A comparison between the performances of target recognition systems that fuse the observed signatures with those that fuse the identification decision is established. Multi-radar systems interrogating the same target provide multi-aspect signatures that can enhance, when fused properly, the recognition performance. This paper proposes a signatures fusion scheme as well as a decision fusion mechanism. The recognition performance assuming a target with unknown azimuth is assessed using both systems, and a comparison is established. The performance is assessed using real radar signatures recorded in a compact range setting. The paper highlights scenarios where a multi-radar system provides little advantage for target recognition purposes, and when it can significantly improve the recognition performance. The same approach can be used when dealing with other types of networked sensors.

We present a novel algorithm for tracking with ladar sensors to aid in navigation, guidance and control systems, suitable for applications to unmanned air vehicles. The methods we employ are based on Bayesian segmentation, optical flow, active contour and Bayesian particle tracking. The algorithm herein holds several significant advantages over traditional tracking methods. The first step in the process is the optimal segmentation of images to enhance the targets and extract them from background clutter and noise. The Bayesian approach to segmentation allows the use of intensity (passive) and range (active) imagery to find targets. Optical flow generalizes and improves correlation techniques for locating objects within a frame, allowing for aspect angle and range changes. With optical flow, we may infer relative velocities on a pixel-by-pixel basis. Active contours are ideally suited to both target-sparse and target-rich environments. The energy approach to determining contours allows the merging and separating of potential targets in an automatic manner. Bayesian particle tracking techniques are used to track the contours over time. The algorithm is tested successfully on experimental and simulated ladar data (using both intensity and range data) as well as sequences of video imageries. The streamlined processing, from obtaining the image data (of size 805x148 pixels) to detecting the moving target to wrapping an active contour on the target, takes less than one second of clock time and provides very accurate predictions of the target location in future frames.

Ensemble methods provide a principled framework in which to build high performance classifiers and represent many types of data. As a result, these methods can be useful for making inferences about biometric and biological events. We introduce a novel ensemble method for combining multiple representations (or views). The method is a multiple view generalization of AdaBoost. Similar to AdaBoost, base classifiers are independently built from each represetation. Unlike AdaBoost, however, all data types share the same sampling distribution computed from the base classifier having the smallest error rate among input sources. As a result, the most consistent data type dominates over time, thereby significantly reducing sensitivity to noise. The method is applied to the problem of facial and gender prediction based on biometric traits. The new method outperforms several competing techniques including kernel-based data fusion, and is provably better than AdaBoost trained on any single type of data.

The Sensors Directorate of the Air Force Research Laboratory (AFRL), in conjunction with the Global Hawk Systems Group, the J-UCAS System Program Office and contractor Defense Research Associates, Inc. (DRA) is conducting an Advanced Technology Demonstration (ATD) of a sense-and-avoid capability with the potential to satisfy the Federal Aviation Administration's (FAA) requirement for Unmanned Aircraft Systems (UAS) to provide "an equivalent level of safety, comparable to see-and-avoid requirements for manned aircraft". This FAA requirement must be satisfied for UAS operations within the national airspace. The Sense-and-Avoid, Phase I (Man-in-the-Loop) and Phase II (Autonomous Maneuver) ATD demonstrated an on-board, wide field of regard, multi-sensor visible imaging system operating in real time and capable of passively detecting approaching aircraft, declaring potential collision threats in a timely manner and alerting the human pilot located in the remote ground control station or autonomously maneuvered the aircraft. Intruder declaration data was collected during the SAA I & II Advanced Technology Demonstration flights conducted during December 2006. A total of 27 collision scenario flights were conducted and analyzed. The average detection range was 6.3 NM and the mean declaration range was 4.3 NM. The number of false alarms per engagement has been reduced to approximately 3 per engagement.

Gradient direction models for corners of prescribed acuteness, leg length, and leg thickness are constructed by generating fields of unit vectors emanating from leg pixels that point normal to the edges. A novel FFT-based algorithm that quickly matches models of corners at all possible positions and orientations in the image to fields of gradient directions for image pixels is described. The signal strength of a corner is discussed in terms of the number of pixels along the edges of a corner in an image, while noise is characterized by the coherence of gradient directions along those edges. The detection-false alarm rate behavior of our corner detector is evaluated empirically by manually constructing maps of corner locations in typical overhead images, and then generating different ROC curves for matches to models of corners with different leg lengths and thicknesses. We then demonstrate how corners found with our detector can be used to quickly and automatically find families of polygons of arbitrary position, size and orientation in overhead images.

There are many advantages of using acoustic sensor arrays to perform targets of interest identification and classification in the battlefield. They are low cost and have relatively low power consumption. They require no line of sight and provide many capabilities for target detection, bearing estimation, target tracking, classification and identification. Furthermore, they can provide cueing for other sensors and multiple acoustic sensors responses can be combined and triangulated to localize an energy source target in the field. In practice, however, many environment noise, time-varying, and uncertainties factors affect their performance in detecting targets of interest reliably and accurate. In this paper, we have proposed a novel feature extraction approach for robust classification and identification of moving target vehicles to reduce those factors. The approach is based on Low Rank Decomposition based L_p norm. Using Low Rank Decomposition based L₁ norm where p = 1, dominant features of vehicle acoustic signatures can be extracted appropriately with respect to vehicle operational responses and used for robust identification and classification of target vehicles. The performance of the proposed approach has been evaluated based on a set of experimental acoustic data from multiple vehicle test-runs. It is demonstrated that the approach yields significant improvement results over our earlier vehicle classification technique based on Singular Value Decomposition (SVD) and reduces uncertainties associated with classification of target vehicles based on acoustic signatures at different operation speeds in the field.

Military Operations in Urban Terrain (MOUT) require the capability to perceive and to analyse the situation around a patrol in order to recognize potential threats. Human operators can only observe a limited field of regard. Sensors can enhance the field of regard up to 360°, but then the amount of data cannot be fully exploited by a human operator any more. For this reason an intelligent assistance system is required that monitors the circumference of a moving platform and warns the driver of a threatening situation. One first processing step of such a system is the recognition of humans. There are numerous approaches to the detection of humans, mainly from stationary cameras. Moving cameras play a role in the field of pedestrian protection from a moving road vehicle. There are two principal differences to this latter application domain. Firstly, the threat in a MOUT scenario potentially arises from humans in the scene. Secondly, not only the trajectories of individual humans are relevant, but also the motion and the behavior of groups of humans. As a first step towards an assistance system that automatically warns drivers in a MOUT scenario, we implemented an approach to the detection of humans in video images and applied them to a relevant set of image sequences taken in a MOUT scenario. In the paper we assess the obtained results and outline further research activities.

Surveillance applications are primarily concerned with detection of targets. In electro-optical surveillance systems, missiles or other weapons coming towards you are observed as moving points. Typically, such moving targets need to be detected in a very short time. One of the problems is that the targets will have a low signal-to-noise ratio with respect to the background, and that the background can be severely cluttered like in an air-to-ground scenario. The first step in detection of point targets is to suppress the background. The novelty of this work is that a super-resolution reconstruction algorithm is used in the background suppression step. It is well-known that super-resolution reconstruction reduces the aliasing in the image. This anti-aliasing is used to model the specific aliasing contribution in the camera image, which results in a better estimate of the clutter in the background. Using super-resolution reconstruction also reduces the temporal noise, thus providing a better signal-to-noise ratio than the camera images. After the background suppression step common detection algorithms such as thresholding or track-before-detect can be used. Experimental results are given which show that the use of super-resolution reconstruction significantly increases the sensitivity of the point target detection. The detection of the point targets is increased by the noise reduction property of the super-resolution reconstruction algorithm. The background suppression is improved by the anti-aliasing.

A new algorithm for pose estimation of vehicles in SAR imagery is presented. Using robust features and a structured decision process, the algorithm achieves high precision. Four neural networks are used to make estimates conditional on angular regions, and another neural network is used to fuse these estimates. For the MSTAR Test Sample, the absolute error has a mean of 2 degrees with a standard deviation of 2.1, which is significantly more precise than previously reported results.

Quadratic Correlation Filters have recently been used for Automatic Target Recognition (ATR). Among these, the Rayleigh Quotient Quadratic Correlation Filter (RQQCF) was found to give excellent performance when tested extensively with Infrared imagery. In the RQQCF method, the filter coefficients are obtained, from a set of training images, such that the response to the filter is large when the input is a target and small when the input is clutter. The method explicitly maximizes a class separation metric to obtain optimal performance. In this paper, a novel transform domain approach is presented for ATR using the RQQCF. The proposed approach, called the Transform Domain RQQCF (TDRQQCF) considerably reduces the computational complexity and storage requirements, by compressing the target and clutter data used in designing the QCF. Since the dimensionality of the data points is reduced, this method also overcomes the common problem of dealing with low rank matrices arising from the lack of large training sets in practice. This is achieved while retaining the high recognition accuracy of the original RQQCF technique. The proposed method is tested using IR imagery, and sample results are presented which confirm its excellent properties.

Image decomposition using directional filter banks is useful in discovering and extracting edge orientation cues for target detection in airborne surveillance images. Since images of interest are very large and the filtered images are not downsampled in the application of interest, conventional filtering can be computationally extremely demanding and there is a need to explore procedures to make the filtering efficient. In this paper a novel filter bank structure for directional filtering of images is proposed and its design described. The design is carried out by imposing structural constraints on the filters, which are implemented using a generalized notion of separable filtering. The structure uses one-dimensional (1-D) filters as building blocks, which are employed in novel configurations to obtain filters with narrow wedge-shaped passbands. Design procedures have been developed for constructing 16-band, 32-band, and 64- band partitions starting with either built-in or user-specified 1-D prototypes. Implementations of filters using the proposed method show significant improvement compared with conventional implementation, often more by an order of magnitude, which is also supported by a theoretical analysis of the filter complexity.

A pose search algorithm is used in conjunction with a bank of binary phase only filters (BPOF's) to determine the pose of a craft in a docking scenario. This approach is facilitated by the use of a high speed spatial light modulator (SLM) capable of operating at ~ 4 kHz integrated into a VanderLugt type optical correlator. The filters are generated using images created from a 3D model of a craft and 3D animation software. The pose estimation is calculated using a computer interfaced to the optical correlator via a Matlab developed algorithm that captures the image frame and performs a series of correlations from a subset of filters selected by the algorithm from the filter bank.

Receiver operator characteristic (ROC) analysis is an emerging automated target recognition system performance assessment tool. The ROC metric, area under the curve (AUC), is a universally accepted measure of classifying accuracy. In the presented approach, the detection algorithm output, i.e., a response plane (RP), must consist of grayscale values wherein a maximum value (e.g. 255) corresponds to highest probability of target locations. AUC computation involves the comparison of the RP and the ground truth to classify RP pixels as true positives (TP), true negatives (TN), false positives (FP), or false negatives (FN). Ideally, the background and all objects other than targets are TN. Historically, evaluation methods have excluded the background, and only a few spoof objects likely to be considered as a hit by detection algorithms were a priori demarcated as TN. This can potentially exaggerate the algorithm's performance. Here, a new ROC approach has been developed that divides the entire image into mutually exclusive target (TP) and background (TN) grid squares with adjustable size. Based on the overlap of the thresholded RP with the TP and TN grids, the FN and FP fractions are computed. Variation of the grid square size can bias the ROC results by artificially altering specificity, so an assessment of relative performance under a constant grid square size is adopted in our approach. A pilot study was performed to assess the method's ability to capture RP changes under three different detection algorithm parameter settings on ten images with different backgrounds and target orientations. An ANOVA-based comparison of the AUCs for the three settings showed a significant difference (p<0.001) at 95% confidence interval.

Automatic target detection (ATD) systems process imagery to detect and locate targets in imagery in support of a variety of military missions. Accurate prediction of ATD performance would assist in system design and trade studies, collection management, and mission planning. A need exists for ATD performance prediction based exclusively on information available from the imagery and its associated metadata. We present a predictor based on image measures quantifying the intrinsic ATD difficulty on an image. The modeling effort consists of two phases: a learning phase, where image measures are computed for a set of test images, the ATD performance is measured, and a prediction model is developed; and a second phase to test and validate performance prediction. The learning phase produces a mapping, valid across various ATR algorithms, which is even applicable when no image truth is available (e.g., when evaluating denied area imagery). The testbed has plug-in capability to allow rapid evaluation of new ATR algorithms. The image measures employed in the model include: statistics derived from a constant false alarm rate (CFAR) processor, the Power Spectrum Signature, and others. We present performance predictors for two trained ATD classifiers, one constructed using using GENIE Pro^TM, a tool developed at Los Alamos National Laboratory, and the other eCognition^TM, developed by Definiens (http://www.definiens.com/products). We present analyses of the two performance predictions, and compare the underlying prediction models. The paper concludes with a discussion of future research.

The identification of a target from an electro-optical or thermal imaging sensor requires accurate sensor registration, quality sensor data, and an exploitation algorithm. Combining the sensor data and exploitation, we are concerned with developing an electro-optical or infrared (EO/IR) performance model. To combat the registration issue, we need a detailed list of operating conditions (i.e. collection position) so that the sensor exploitation results can be evaluated with sensitivities to these operating conditions or collection parameters. The focus of this paper will build on the NVSED AQUIRE model². We are also concerned with developing an EO/IR model that affords comparable operating condition parameters to a synthetic aperture radar (SAR) performance model. The choice of EO/IR modeling additions are focused on areas were Fusion Gain might be realized through an experiment tradeoff between multiple EO/IR looks for ATR exploitation fusion. The two additions to known EO/IR models discussed in the paper are (1) adjacency and (2) obscuration. The methods to account for these new operating conditions and the corresponding results on the modeled performance are presented in this paper.

Many aspects of detonation phenomena have been well studied over the last century. However, the transient infrared and visible emissions from detonation fireballs have been poorly understood, and this has hampered attempts to remotely identify explosives via combustion signatures. Recently, time-resolved infrared spectra (1800-7000 cm^-1, 4cm^-1 resolution, 8 Hz) were collected from the detonation of uncased charges of TNT and several kinds of improvised explosive devices in four weight classes (10, 50, 100, and 1000 kg). A simple model for fireball emissions has been developed which accurately describes the observed spectra in terms of the fireball size, temperature, gaseous byproduct concentrations, and grey particulate absorption coefficient. The model affords high-fidelity dimensionality reduction and provides physical features which can be used to distinguish the uncased explosives.

A variety of change detection (CD) methods have been developed and employed to support imagery analysis for applications including environmental monitoring, mapping, and support to military operations. Evaluation of these methods is necessary to assess technology maturity, identify areas for improvement, and support transition to operations. This paper presents a methodology for conducting this type of evaluation, discusses the challenges, and illustrates the techniques. The evaluation of object-level change detection methods is more complicated than for automated techniques for processing a single image. We explore algorithm performance assessments, emphasizing the definition of the operating conditions (sensor, target, and environmental factors) and the development of measures of performance. Specific challenges include image registration; occlusion due to foliage, cultural clutter and terrain masking; diurnal differences; and differences in viewing geometry. Careful planning, sound experimental design, and access to suitable imagery with image truth and metadata are critical.

We present a simple formula relating any two probability densities and show that this relation generalizes a number of results of classical probability theory, such as the Gram-Charlier and Edgeworth series. Our generalization is twofold in that it relates two arbitrary distributions rather than one of them being a Gaussian and secondly we generalize to an arbitrary Hermitian operator rather than the differentiation operator.

Human motion tracking is an active area of research in computer vision and machine intelligence. It has many applications in video surveillance and human-computer interface. Most of the existing algorithms track multiple humans in a given image. This paper proposes a detection approach which can track a specific person from a crowded environment. Mean shift clustering algorithm is employed in the difference image to get the candidate cluster which is found to converge within few iterations. The number of clusters and the cluster centers are automatically derived by mode seeking with the mean shift procedure. Discrete cosine transform is applied to each cluster and to the known target to extract features of the clusters and the target. To get the target cluster from a given image, Mahalanobis distance is measured between each transformed candidate cluster and the target. The cluster with the minimum distance is taken as the desired target. Tracking is carried out by updating the cluster parameters over time using the mean shift procedure.

The large number of rear end collisions due to driver inattention has been identified as a major automotive safety issue. Even a short advance warning can significantly reduce the number and severity of the collisions. In this paper, we describe an image alignment based vehicle tracking method that employs only a single moving camera mounted on the driver's automobile as input, for use in detecting rear vehicles on highways and city streets. We also present a method to compute the relative distances between the rear vehicles and the driver's car. With the aid of symmetrical function and simplified image alignment tracking methodology, our methodology becomes relatively simple to implement using embedded system technology in the automobile environment with real-time multiple vehicles tracking and successful rate over 97%.

We overview a technique known as Super Image Vector Inner Product as applied to Facial pose estimation. The method is mathematically similar to correlation based methods but is numerically more efficient. The Vector Inner Product approach attains its pose and position estimation by embedding these distortions in its phase response. We demonstrate for the first time, that that the Super Image Vector Inner Product can be used for facial identification. We present a method by which segmenting the face into a set of feature regions, individual Super Images can be combined together using mesh techniques to track facial expressions.

Automatic Target Recognition (ATR) using three-dimensional (3D) sensor data has proven very successful in experimental platforms. One of the factors limiting the implementation of these approaches is lag in operational hardware to provide the type of data required. Neptec has addressed this sensor concern in its 3D ATR software. The need for specific operational 3D sensing hardware is avoided by using a generic range image format, and a shape-from-motion (SfM) method enables the generation of 3D data using widely available 2D sensors. The previously reported ATR software has been expanded from proof-of-concept ground-to-ground to include air-to-ground capabilities. The system uses a generic 3D model of the target, such as from CAD or scanned from a scale or full-sized model which does not need to be perfect. The rapid recognition approach simultaneously provides target pose estimation. This capability has been demonstrated using ground-based imaging LiDAR, airborne LiDAR, scannerless AMcw LiDAR, and shape-from-motion using a 2D camera. Multiple data sets can be fused to optimize confidence in the recognition and provide measures of similarity between different targets and the data set. This paper presents an overview of the 3D ATR approach and updates performance characteristics from a variety of tests that include synthetic data, lab tests, and field tests. It is shown that the approach is fast, highly robust, flexible, and is primarily limited by the quality of sensor data. Particular emphasis is placed on the shape-from-motion application since this capability can make use of widely used operational 2D imaging sensor packages.

This paper illustrates a statistical model-based approach to the problem of target detection in a cluttered scene from long-wave infrared images, accommodating both unknown range to the target, unknown target location in the image, and unknown gain control settings on the imaging device. The philosophical perspective adopted emphasizes an iterative process of model creation and refinement and subsequent evaluation. The overarching theme is on the clear statement of all assumptions regarding the relationships between ground truth and corresponding imagery, the assurance that each admits quantifiable refutation, and the opportunity costs associated with their adoption for a particular problem.