This paper describes a model-based 3D object recognition system, which makes use of 3D data acquired by LIDAR sensors. The system is based on a coarse-to-fine scheme for object indexing and verification to achieve high efficiency and accuracy. The system employs rotationally invariant semi-local spin image features for object representation and formulates the recognition process as a feature searching through a database, followed by a matching process guided by putative candidates between the query features and the model features. To achieve recognition efficiency with sublinear dependency on the size of the model database, an approximate nearest-neighbor method, the locality-sensitivity-hashing (LSH), is used for feature search. Geometrically constrained scene-model correspondences are used to generate alignment hypotheses that are refined during a matching and verification process for achieving high recognition accuracy. A large model database of commercial and military vehicles is used for experiments. Results on real data acquired with commercial LADAR sensor systems, mounted on either high-lift or airborne platforms are presented. Our results indicate that 3D object recognition on LADAR data is matured to the point that it is ready for real large-scale applications.

In this work we investigate simultaneous object identification improvement and efficient library search for model-based object recognition applications. We develop an algorithm to provide efficient, prioritized, hierarchical searching of the object model database. A common approach to model-based object recognition chooses the object label corresponding to the best match score. However, due to corrupting effects the best match score does not always correspond to the correct object model. To address this problem, we propose a search strategy which exploits information contained in a number of representative elements of the library to drill down to a small class with high probability of containing the object. We first optimally partition the library into a hierarchic taxonomy of disjoint classes. A small number of representative elements are used to characterize each object model class. At each hierarchy level, the observed object is matched against the representative elements of each class to generate score sets. A hypothesis testing problem, using a distribution-free statistical test, is defined on the score sets and used to choose the appropriate class for a prioritized search. We conduct a probabilistic analysis of the computational cost savings, and provide a formula measuring the computational advantage of the proposed approach. We generate numerical results using match scores derived from matching highly-detailed CAD models of civilian ground vehicles used in 3-D LADAR ATR. We present numerical results showing effects on classification performance of significance level and representative element number in the score set hypothesis testing problem.

Land surface elevation measurements from airborne laser swath mapping (ALSM) data can be irregularly spaced due to occlusion by forest canopy or scanner and aircraft motion. The measurements are usually interpolated into a regularly spaced grid using techniques such as Kriging or spline-interpolation. In this paper a probabilistic graphical model called a Bayesian network (BN) is employed to interpolate missing data. A grid of nodes is imposed over ALSM measurements and the elevation information at each node is estimated using two methods: 1) a simple causal method, similar to a Markov mesh random field (MMRF), and 2) BN belief propagation. The interpolated results of both algorithms using the maximum a posteriori (MAP) estimates are presented and compared. Finally, uncertainty measures are introduced and evaluated against the final estimates from the BN belief propagation algorithm.

In this keynote address, we introduce three-dimensional (3D) sensing, visualization and recognition of microorganisms using microscopy-based single-exposure on-line (SEOL) digital holography. A coherent Mach-Zehnder interferometer records Fresnel diffraction field by a single on-line exposure to generate a microscopic digital hologram. Complex amplitude distribution is numerically reconstructed by the inverse Fresnel transform at arbitrary depth planes. After the reconstruction of volumetric complex images, 3D biological micro-objects are segmented and features are extracted by Gabor-based wavelets. The graph matching technique searches predefined 3D morphological shapes of reference biological microorganisms. Preliminary experimental results using sphacelaria alga and tribonema aequale alga are presented.

Shape measurements form powerful features for recognizing objects, and many imaging modalities produce three-dimensional shape information. Stereo-photogrammetric techniques have been extensively developed, and many researchers have looked at related techniques such as shape from motion, shape from accommodation, and shape from shading. Recently, considerable attention has focused on laser radar systems for imaging distant objects, such as automobiles from an airborne platform, and on laser-based active stereo imaging for close-range objects, such as part scanners for automated inspection. Each use of these laser imagers generally results in a range image, an array of distance measurements as a function of direction. For multi-look data or data fused from multiple sensors, we may more generally treat the data as a 3D point-cloud, an unordered collection of 3D points measured from the surface of the scene. This paper presents a general approach to object recognition in the presence of significant clutter, that is suitable for application to a wide range of 3D imaging systems. The approach relies on a probabilistic framework relating 3D point-cloud data and the objects from which they are measured. Through this framework a minimum probability of error recognition algorithm is derived that accounts for both obscuring and nonobscuring clutter, and that accommodates arbitrary (range and cross-range) measurement errors. The algorithm is applied to a problem of target recognition from actual 3D point-cloud data measured in the laboratory from scale models of civilian automobiles. Noisy 3D measurements are used to train models of the automobiles, and these models are used to classify the automobiles when present in a scene containing natural and man-made clutter.

Extraction and efficient representation of informative structure from data is the goal of pattern recognition. Efficient and effective parametric and nonparametric representations for capturing the geometry of three-dimensional objects are an area of current research. Tang and Medioni have proposed tensor representations for characterization and reconstruction of surfaces. 3-D structure tensors are extracted by mapping surface geometries using a rank-2 covariant tensor. Distributional differences between representations of objects of interest can (theoretically) be used for target matching and identification. This paper analyzes the statistical distributions of tensor representation extracted from 3-D LADAR imagery and quantifies a measure of divergence between images of three vehicles as a function of tensor feature support size.

In this paper, we describe and implement a genetic algorithm based composite wavelet-matched filter for target recognition. The designed filter is invariant to 0-to-360° out-of-plane rotation ranges. The Mexican hat wavelet has been used for the design of the filter. Genetic algorithm has been used to optimize the weight factors for the input matched filters and the different scale wavelets. The designed filter has been implemented in the hybrid digital-optical correlator architecture. Simulation and experimental results in support of the proposed idea are presented.

Modern computer vision systems lack human-like abilities to understand the visual scene, detect, and unambiguously identify and recognize objects. Bottom-up grouping can rarely be effective for real world images if applied to the whole image without having clear criteria of how to further combine obtained small distinctive neighbor regions into meaningful objects. ATR systems that are based on the similar principles become dysfunctional if a target doesn't demonstrate remarkably distinctive and contrasting features that allow for unambiguous separation from background and identification. However, human vision unambiguously separates any object from its background and recognizes it, using a rough but wide peripheral system that tracks motions and regions of interests, and narrow but precise foveal vision that analyzes and recognizes the object in the center of a selected region of interest, and visual intelligence that provides scene and object contexts and resolves ambiguity and uncertainty in the visual information. Biologically-inspired Network-Symbolic models convert image information into an "understandable" Network-Symbolic format, which is similar to relational knowledge models. The equivalent of interaction between peripheral and foveal systems in the network-symbolic system is achieved via interaction between Visual and Object Buffers and top-level knowledge system. This interaction provides recursive rough context identification of regions of interest in the visual scene and their analysis in the object buffer for precise and unambiguous separation of the target from clutter with following the recognition of the target.

Estimation of a remote object's physical characteristics, such as size, shape and optical cross-section (OCS) can provide valuable strategic and tactical information. Imaging systems, such as those employing adaptive optics, can provide excellent images of space objects through a corrupting atmosphere. However, such systems are extremely expensive when thought of from the viewpoint of queuing sensors. A queuing sensor may interrogate a target of interest on a regular basis and if substantial changes in target characteristics are discovered, the sensor can alert an imaging system to investigate. The techniques discussed in this paper use only the total received time-series signal from a laser illumination experiment. The authors have specialized in the analysis of such signals and have shown that estimates of laser pointing disruptions, known as jitter and boresight, may be made using χ² tests. Moreover, these estimates, as well as others, may be performed in real time, far more useful than post-processing. Nukove is currently supported by an AFOSR Phase II STTR to develop a prototype software tool to provide realtime estimates and the technique has been demonstrated successfully in a laboratory environment. This paper studies the potential for a well-controlled system to determine approximate target size and shape using statistical χ² techniques similar to the pointing techniques. Moreover, such estimates may be made in realtime. Since the data is taken from a full aperture and not a focal plane, effects such as speckle and scintillation, which corrupt imaging systems, have been shown to have little impact when the receiving aperture is on the order of one meter or larger.

This paper presents a novel methodology for target recognition and reconstruction of rigid body moving targets. Traditional methods such as Synthetic Aperture Radar (SAR) rely on information gathered from multiple sensor locations and complex processing algorithms. Additional processing is often required to mitigate the effects of motion and improve the image resolution. Many of these techniques rely on information external to the target such as target radar signatures and neglect information available from the structure of the target, structural invariance, and kinematics. This revolutionary target reconstruction method incorporates information not traditionally used. As a result, the absolute position of target scattering centers can theoretically be determined with external, high resolution radar range information from three observations of four target scattering centers. Relative motion between the target and the sensor and structural invariance provide additional information for determining position of the target's scattering center, actual scaling, and angular orientation with respect to the sensor for reconstruction and imaging. This methodology is based on the kinematics of rotational motion resulting from relative movement between the sensor and the target. External range data provides one-dimensional information for determining position in a two-dimensional projection of the scattering centers. The range location of the scattering center, relative to a defined center, is analyzed using rotational motion. Range and target kinematics support the development of a conceptual model. Actual scaling and the target's orientation with respect to the sensor are developed through a series of trigonometric relationships. The resulting three-dimensional coordinates for the scattering centers are then used for target reconstruction and image enhancement.

In this paper, we proposed a three dimensional matching algorithm using geometrical invariants. Invariant relations between 3D objects and 2D images for object recognition has been already developed in Ref. [1]. We proposed a geometrical invariant approach for finding relation between 3D model and stereo image pair. Since the depth information is lost in a single 2D image, we cannot recognize an object perfectly. By constructing a 3D invariant space we can represent a 3D model as a set of points in the invariant space. While matching with the 2D image we can draw a set of invariant light rays in 3D, each ray passing through a 3D invariant model point. If enough rays intersect the model in 3D invariant space we can assume that the model is present in the image. But for a single image the method is not that much reliable as the depth information is never considered. In the proposed method, as the matching is performed using stereo image pair, it is more reliable and accurate.

The Modified Eigenvalue problem arises in many applications such as Array Processing, Automatic Target Recognition (ATR), etc. These applications usually involve the Eigenvalue Decomposition (EVD) of matrices that are time varying. It is desirable to have methods that eliminate the need to perform an EVD every time the matrix changes but instead update the EVD adaptively, starting from the initial EVD. In this paper, we propose a novel Optimal Adaptive Algorithm for the Modified EVD problem (OAMEVD). Sample results are presented for an ATR application, which uses Rayleigh Quotient Quadratic Correlation filters (RQQCF). Using a Infrared (IR) dataset, the effectiveness of this new technique as well as its advantages are illustrated.

This study is an initial investigation into the efficacy of texture operators for detection of military vehicle targets-in-the-clear in SAR imagery. The specific study is a very simple problem that aims to evaluate a particular feature set that arises in an approach to computer vision called spatial spectroscopy. Spatial spectroscopy begins by partitioning the image's spatial (Fourier) spectrum using a bank of filters. The filters compute a multiscale, truncated Taylor Series expansion at each pixel. Suitably extended on generic images, this feature space is capable of producing a unique pattern describing each pixel. The objective, of course, is not to uniquely distinguish each pixel but to form groups of pixels corresponding to targets in SAR that are distinct from background pixels. Thus, nonlinear operators are required to fold, twist, and bend the feature space in ways that cause pixels that make up targets to group together. The particular nonlinear operators for a study depend on the invariances and equivariances of the problem. In the present case, a large suite of operators is applied to the image data and principal discriminant analysis is used to select the most relevant features. Texture operators are found to be effective at discriminating targets from background.

Often sensor ego-motion or fast target movement causes the target to temporarily go out of the field-of-view leading to reappearing target detection problem in target tracking applications. Since the target goes out of the current frame and reenters at a later frame, the reentering location and variations in rotation, scale, and other 3D orientations of the target are not known thus complicating the detection algorithm has been developed using Fukunaga-Koontz Transform (FKT) and distance classifier correlation filter (DCCF). The detection algorithm uses target and background information, extracted from training samples, to detect possible candidate target images. The detected candidate target images are then introduced into the second algorithm, DCCF, called clutter rejection module, to determine the target coordinates are detected and tracking algorithm is initiated. The performance of the proposed FKT-DCCF based target detection algorithm has been tested using real-world forward looking infrared (FLIR) video sequences.

The use of 3-Dimensional information in face recognition requires pose estimation. We present the use of 3-Dimensional composite correlation filter to obtain pose estimation without the need for feature identification. Composite correlation filter research has been vigorously pursued in the last three decades due to their applications in many areas, but mainly in distortion-invariant pattern recognition. While most of this research is in two-dimensional space, we have extended our study of composite filters to three-dimensions, specifically emphasizing Linear Phase Coefficient Composite Filter (LPCCF). Unlike previous approaches to composite filter design, this method considers the filter design and the training set selection simultaneously. In this research, we demonstrate the potential of implementing LPCCF in head pose estimation. We introduce the utilization of LPCCF in the application of head pose recovery through full correlation using a set of 3-D voxel maps instead of the typical 2-D pixel images/silhouettes. Unlike some existing approaches to pose estimation, we are able to acquire 3-D head pose without locating salient features of a subject. In theory, the correlation phase response contains information about the angle of head rotation of the subject. Pose estimation experiments are conducted for two degrees of freedom in rotation, that is, yaw and pitch angles. The results obtained are very much inline with our theoretical hypothesis on head orientation estimation.

The joint transform correlator (JTC) is one of two main optical image processing architectures which provide us with a highly effective way of comparing images in a wide range of applications. Traditionally an optical correlator is used to compare an unknown input scene with a pre-captured reference image library, to detect if the reference occurs within the input. There is a new class of application for the JTC where they are used as image comparators, not having a known reference image, rather frames from a video sequence form both the input and reference. The JTC input plane is formed by combining the current frame with the previous frame in a video sequence and if the frames match, then there will be a correlation peak. If the objects move then the peaks will move (tracking) and if something has changed in the scene, then the correlation between the two frames is lost. This forms the basis of a very powerful application for the JTC in Defense and Security. Any change in the scene can be recorded and with the inherent shift invariance property of the correlator, any movement of the objects in the scene can also be detected. A major limitation of the JTC is its intolerance to rotation and scale changes in input compared to the reference images. The strength of the correlation signal decreases as the input object rotates or varies in scale relative to the reference object. We have designed binary phase only filters using the direct binary search algorithm for rotation invariant pattern recognition for a 1/f JTC. Simulation and experimental results are included. If the relative alignment of the images in the input plane is known then the desirable fringes in the resulting joint power spectrum (JPS) can be selectively enhanced during the binarisation process. This can have a highly beneficial effect on the resulting correlation intensities. For the input plane in which input and reference images are placed side by side we develop the vertical edge enhancement (VEE) technique that concentrate solely on the vertical components of the JPS during the binarisation process. Simulation and experiments proves that VEE enhances the correlation intensities and suppresses the zero order noise.

In active sonar or radar, the received signal can often be modeled as a convolution of the transmitted signal with the channel impulse response and the target impulse response. Because the received signal may have a time-varying spectrum, due for example to target motion or to changes in the channel impulse response, time-frequency methods have been used to characterize propagation effects and target effects, and to extract features for classification. In this paper, we consider the time-varying spectrum, in particular the Wigner time-frequency representation, of a received signal modeled as the convolution of the transmitted signal with the channel and target responses. We derive a simple but insightful approximation that shows the effects of the magnitude and phase of the frequency response of the target and of the channel on the Wigner representation of the transmitted signal. We also consider time-varying effects on the Wigner representation, such as changes in reflected energy, which we model by amplitude modulation.

A lot of efforts have been pursued in Automatic Target Recognition (ATR), including, based on: Fourier transform, wavelet transform, novelty filtering, and many others. Unfortunately, in all these methods, a target, either on-the-move (OTM), or static, has to already be identified. Such a Target Identification (ID) pre-ATR process, however, requires significant data reduction that must be done before the ATR process starts. The pre-ATR ID process becomes complex if it is performed by RF-wireless visual sensor networks, based on Unmanned Ground Vehicles (UGVs), or other ground vehicles. The visual sensors include: TV cameras, artificial animal eyes [1] (fish eye, bug eye, lobster eye [2], etc.), and other video-like sensors. These sensors need to work not only autonomously autonomously, but also in cooperation , through RFwireless inter-communication which should be continuous to preserve constant cooperation. Such constant video communication should avoid video image breakdown (in the form of heavy pixeling, or complete image blackout) under abrupt worsening of digital data transfer conditions, in terms of increasing environmental noise (or, reducing SNR), and/or increasing of BER (bit-error-rate) of the video signal transfer. Thus, replacement of image breakdown by its graceful degradation (i.e., preserving image continuity at the expense of quality reduction) is a central issue of the realtime pre-ATR video data reduction. In this paper, we will first discuss the conditions of continuous video RF-communication, based on graceful image degradation, and then analyze some important examples of the real-time pre-ATR video data reduction, based on cooperative video networks, with TV-cameras, as visual sensors.

This paper presents a system which automatically detects moving targets contained in aerial sequences of FLIR images under heavy cluttered conditions. In these situations, the detection of moving targets is generally carried out through the implementation of segmentation and tracking techniques based on the images correlation maintained by the static camera hypothesis. However, detection procedures cannot rely on this correlation when the camera is airborne and, therefore, image stabilization techniques are usually introduced previously to the detection process. Nevertheless, the use of stabilization algorithms has been often applied to terrestrial sequences and assuming a high computational cost. To overcome these limitations, we propose an innovative and efficient strategy, with a block-based estimation and an affine transformation, operating on a multi-resolution approach for recovering from the ego-motion. Next, once the images have been compensated on the highest resolution image and refined to avoid distortions produced in the sampling process, a dynamic differences-based segmentation followed by a morphological filtering strategy is applied. The novelty of our strategy relies on the relaxation of the pre-assumed hypothesis and, hence, on the enhancement of its applicability, and also by further reducing its computational cost, thanks to the application of a multi-resolution algorithm. The experiments performed have obtained excellent results and, although the complexity of the system arises, the application of the multi-resolution approach has proved to dramatically reduce the global computational cost.

In this paper, we propose a target detection algorithm in FLIR imagery using independent component analysis (ICA). Here FLIR images of some real targets with practical background regions are used for training. Dimension of the training regions is chosen depending on the size of the target. After performing ICA transformation on these training images, we obtain a ICA matrix, where each row gives the transformed version of the previous matrix, and a weight matrix. Using these matrices, a transformed matrix of the input image can be found with enhanced features. Then cosine of the angle between the training and test vectors is employed as the parameter for detecting the unknown target. A test region is selected from the first frame of FLIR image, which is of the same size as the training region. This region is transformed following the proposed algorithm and then the cosine value is measured between this transformed vector and the corresponding vector of the transformed training matrix. Next the test region is shifted by one pixel and the same transformation and measurement are done. Thus the whole input frame is scanned and we get a matrix for cosine values. Finally a target is detected in a region of the input frame where it gives the highest cosine value. A detailed computer simulation program is developed for the proposed algorithm and a satisfactory performance is observed when tested with real FLIR images.

Advances in sensor technology and image processing have made it possible to equip unmanned aerial vehicles (UAVs) with economical, high-resolution, energy-efficient sensors. Despite the improvements, current UAVs lack autonomous and collaborative operation capabilities, due to limited bandwidth and limited on-board image processing abilities. The situation, however, is changing. In the next generation of UAVs, much image processing can be carried out onboard and communication bandwidth problem will improve. More importantly, with more processing power, collaborative operations among a team of autonomous UAVs can provide more intelligent event detection capabilities. In this paper, we present ideas for developing a system enabling target recognitions by collaborative operations of autonomous UAVs. UAVs are configured in three stages: manufacturing, mission planning, and deployment. Different sets of information are needed at different stages, and the resulting outcome is an optimized event detection code deployed onto a UAV. The envisioned system architecture and the contemplated methodology, together with problems to be addressed, are presented.

Techniques for automatic target recognition (ATR) in forward-looking infrared (FLIR) data based on Grenander's pattern theory are revisited. The goal of this work is to unify two techniques: one for multi-target detection and recognition of pose and target type, and another for structured inference of forward-looking infrared (FLIR) thermal states of complex objects. The multi-target detection/recognition task is accomplished through a Metropolis-Hastings jump-diffusion process that iteratively samples a Bayesian posterior distribution representing the desired parameters of interest in the FLIR imagery. The inference of the targets' thermal states is accomplished through an expansion in terms of "eigentanks" derived from a principle component analysis over target surfaces. These two techniques help capture much of the variability inherent in FLIR data. Coupled with future work on rapid detection and penalization strategies to reduce false alarms, we strive for a unified technique for FLIR ATR following the pattern-theoretic philosophy that may be implemented for practical applications.

This paper presents the status of our SAR automatic target recognition (ATR) work on the Moving and Stationary Target Acquisition and Recognition (MSTAR) public database using the minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF). We use a subset of the MSTAR public database for the benchmark three-class problem and we address confuser and clutter rejection. To handle the full 360° range of aspect view in MSTAR data, we use a set of Minace filters for each object; each filter should recognize the object (and its variants) in some angular range. We use fewer DIFs per object than prior work did. The Minace parameter c trades-off distortion-tolerance (recognition) versus discrimination (confuser/clutter rejection) performance. Our filter synthesis algorithm automatically selects the Minace filter parameter c and selects the training set images to be included in the filter, so that the filter can achieve both good recognition and good confuser and clutter rejection performance; this is achieved using a training and validation set. In our new filter synthesis method, no confuser, clutter, or test set data are used. The peak-to-correlation energy (PCE) ratio is used as the correlation plane metric in both filter synthesis and in tests, since it works better than correlation peak height. In tests, we do not assume that the test input's pose is known (as most prior work does), since pose estimation of SAR objects has a large margin of error; we describe our procedure for proper use of pose estimates in MSTAR recognition. The use of circular versus linear correlations is addressed. We also address the use of multi-look SAR data to improve performance.

Optical fiber recirculation loops will change the technical foundation of radar and electronic warfare technologies. It becomes possible to measure Doppler beating with a single pulse, to map out micro Doppler signature with a resolution better than 1.0 Hz, and to take sharp IASR images of targets which are more than several hundred miles away. With fine micro Doppler signature and high precision ISAR images, the passive identification of targets will become a reality.

Target recognition is desirable feature of any defence radar system. With the present revival of interest in bistatic and multi-static radar systems, the future radar systems are predicted to invariably have bistatic abilities. The present project aims at looking into prospects and limitations of bistatic target recognition and to develop efficient algorithm for the same. The work in this paper reports the development of a database of bistatic target signatures, and the application of principal component analysis (PCA) based classifiers on the same. Results are compared with the more basic conditional Gaussian model based Bayesian classifier.

Ground Moving Target Indication (GMTI) radar provides a day/night, all-weather, wide-area surveillance capability to detect moving vehicles and personnel. Current GMTI radar sensors are limited to only detecting and tracking targets. The exploitation of GMTI data would be greatly enhanced by a capability to recognize accurately the detections as significant classes of target. Doppler classification exploits the differential internal motion of targets, e.g. due to the tracks, limbs and rotors. Recently, the QinetiQ Bayesian Doppler classifier has been extended to include a helicopter class in addition to wheeled, tracked and personnel classes. This paper presents the performance for these four classes using a traditional low-resolution GMTI surveillance waveform with an experimental radar system. We have determined the utility of an "unknown output decision" for enhancing the accuracy of the declared target classes. A confidence method has been derived, using a threshold of the difference in certainties, to assign uncertain classifications into an "unknown class". The trade-off between fraction of targets declared and accuracy of the classifier has been measured. To determine the operating envelope of a Doppler classification algorithm requires a detailed understanding of the Signal-to-Noise Ratio (SNR) performance of the algorithm. In this study the SNR dependence of the QinetiQ classifier has been determined.

This paper summarizes the results of experiments in developing Support Vector Machines for polarimetric radar target classification. Previous studies have shown that proper selection of state of polarization in both transmitting and receiving stages can noticeably improve target classification performace. Polarization syntheses is used to generate radar signatures of several targets at various transmit/receive pairs of polarization angles. Then statistical attributes from each radar signature are used for its reperesentation. To address the target separation ambiguities, support vector machines using a number of kernels are developed and used. The results of applying this approach on real fully polarimetric radar data indicate that only a small subset of polarization angles are sufficient for generating signatures needed for training a classifier for optimal separation of targets.

It is known that high frequency in the teraHertz range (THz) have unique advantage in the reflective properties for biological objects and penetration through dielectric material. These radiations are non-ionizing for use in civilian applications. High frequency aperture size can be fairly small allowing the equipment to be portable. THz components mainly consist of sources detectors up-converters down-converters mixers and circulators and associated femto second laser generators. These components are under active development. However each component of these high frequency modules for transmission and reception can be fairly large. In this paper a deterministic thinning procedure is derived for designing an array, with sidelobe control, of these transmission and receiver modules. Circular as well as elliptical arrays are discussed. Algorithm is developed based on Taylor synthesis procedure with zero sampling. Grid locations of these large arrays are given with some examples. Using the results of thinned circular array we design elliptical array using invariant principal of the synthesis. The array design is based on analytic solutions of aperture integral equations. Side lobe control is achieved by controlling the illumination of the aperture. This illumination corresponds to the density of the elements in the sparse array, with each element of the array having uniform amplitude.

Ultra-wideband ground penetrating radar (GPR) systems have proved useful for extracting and displaying information for target recognition purposes. Target signatures whether in the time, frequency, or joint time-frequency domains, will substantially depend on the target's burial conditions such as the type of soil, burial depth, and the soil's moisture content. That dependence can be utilized for target recognition purposes as we have demonstrated previously. The signature template of each target was computed in the time-frequency domain from the returned echo when the target was buried at a known depth in the soil with a known moisture content. Then, for any returned echo the relative difference between the similarly computed target signature and a selected signature template was computed. A global optimization method together with our (approximate) target translation method (TTM) that signature difference, chosen as object function, was minimized by adjusting the depth and moisture content, now taken to be unknown parameters. The template that gave the smallest value of the minimized object function for the returned echo was taken as target classification and the corresponding values of the depth and moisture parameters as estimates of the target's burial conditions. This optimization technique can also be applied to time-series data, avoiding the need for time-frequency analysis. It is then of interest to evaluate the relative merits of time data and time-frequency data for target recognition. Such a comparison is here preformed using signals returned from dummy mines buried underground. The results of the analysis serve to assess the intrinsic worth of data in the time domain and in the time-frequency domain for identifying subsurface targets using a GPR. The targets are buried in a test field at the Swedish Explosive Ordnance Disposal and Demining Center (SWEDEC) at Eksjo, Sweden.

Performance assessment and optimization of ATR systems poses the problem of developing image databases for learning and testing purposes. An automatic IR image database generation technique is presented in this paper. The principle consists in superimposing segmented background, target and mask (bushes for example) from real images, under the constraint of predefined image characterization metrics. Each image is automatically computed according to a specification which defines the metrics levels to reach, such as the local contrast ΔT_RSS (NVESD metric), the Signal to Clutter Ratio, or the masking ratio target/mask. An integrated calibrated sensor model simulates the sensor degradations by using the pre and post-filter MTF, and the 3D noise parameters of the camera. The image generation comes with the construction of a ground truth file which indicates all the parameter values defining the image scenario. A large quantity of images can be generated accordingly, leading to a meaningful statistical evaluation. A key feature is that this technique allows to build learning and testing databases with comparable difficulty, in the sense of the chosen image metrics. The theoretical interest of this technique is presented in the paper, compared to the classical ones which use real or simulated data. An application is also presented, within the CALADIOM project (terrestrial target detection with programmable artificial IR retina combined with IR ATR system). Over 38,000 images were processed by this ATR for training and testing, involving seven armored vehicles as targets.

Automated target recognition (ATR) methods hold promise for rapid extraction of critical information from imagery data to support military missions. Development of ATR tools generally requires large amounts of imagery data to develop and test algorithms. Deployment of operational ATR systems requires performance validation using operationally relevant imagery. For early algorithm development, however, restrictions on access to such data is a significant impediment, especially for the academic research community. To address this limitation, we have developed a set of grayscale imagery as a surrogate for panchromatic imagery that would be acquired from airborne sensors. This surrogate data set consists of imagery of ground order of battle (GOB) targets in an arid environment. The data set was developed by imaging scale models of these targets set in a scale model background. The imagery spans a range of operating conditions and provides a useful image set for initial explorations of new approaches for ATR development.

A novel approach to motion detection in an image sequence is presented. This new approach computes a one-dimensional version of the fractal error metric applied temporally across each pixel. The original fractal error algorithm was developed by Cooper et al. as a two-dimensional metric for detecting man-made features in a single image using only spatial information. The fractal error metric is based on the observed propensity of natural image features to fit a fractional Brownian motion (fBm) model well, thus producing a small fractal error. On the other hand, man-made features do not fit the fBm model well and therefore produce a larger fractal error. Jansing et al. showed that spatial edges typically do not fit the fBm model due to their irregularity. The one-dimensional implementation of the algorithm presented in this paper exploits the irregularity of edges in a temporal signal, which are typically caused by moving objects. Emphasis is placed on moving target detection in the presence of noise and clutter-induced motion. Results are demonstrated using mid-wave infrared (MWIR) image sequences.

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 a performance predictor using a trained classifier ATD that was constructed using GENIE, a tool developed at Los Alamos National Laboratory. The paper concludes with a discussion of future research.

This document describes a method for automatically detecting moving targets in a sequence of wide area scan (WAS) infrared images, where the images may have non-linear geometric distortion from one frame to the next, the time between frames may be large (> 1 second), and the inter-frame motion of the targets may be large relative to the target size. The registration step aligns the images and overcomes the non-linear frame-to-frame geometric distortions. The change detection step detects moving objects in the difference image formed between pairs of registered frames. The final tracking step reduces false alarms and finds moving objects having positive or negative contrast even when the time between frames is large and the target motion between frames is large.

The vibrations produced by objects, for example by a plate or cylinder insonified by a sonar wave, exhibit characteristics unique to the particular structure, which can be used to distinguish among different objects. The situation is complicated, however, by many factors, a particularly important one being propagation through media. As a vibration propagates, its characteristics can change simply due to the propagation channel; for example, in a dispersive channel, the duration of the vibration will increase with propagation distance. These channel effects are clearly detrimental to automatic recognition because they do not represent the object of interest and they increase the variability of the measured responses, especially if measurements are obtained from targets at different locations. Our principal aim is to identify characteristics of propagating vibrations and waves that may be used as features for classification. We discuss various moment-like features of a propagating vibration. In the first set of moments, namely temporal moments such as mean and duration at a given location, we give explicit formulations that quantify the effects of dispersion. Accordingly, one can then compensate for the effects of dispersion on these moments. We then consider another new class of moments, which are invariant to dispersion and hence may be useful as features for dispersive propagation. We present classification results comparing these invariant features to related non-invariant features, for classification of simulated backscatter from different steel shells in a dispersive environment.

Many man made and natural noises in nature are nonstationary, however most methods that have been devised to study noises, theoretically or experimentally, have been devised for the stationary situations. We have developed a number of methods to study noises that are nonstationary. We use the Wigner spectrum and other time-frequency representations to represent time-varying noises. These representations can be thought of as a generalization of the standard power spectrum. In this paper we study simple models of nonstationary noises and obtain the Wigner spectrum numerically from realizations of the noises and also by direct calculation. We show that for our test cases the Wigner spectrum clearly shows the nonstationarities of the noise. We also present a method to generate nonstationary white noise that has very different behavior than the standard white Gaussian noise.

Human body registration is an important and complex problem that can be found in a variety of real world applications. Registration maps images of a person obtained from different camera views into a common reference system of a scene that contains a human figure. The complexity of the problem stems from the fact that human body can arbitrarily move in the 3-D space while changing its own shape. The registration task is stated as a nonlinear global multimodal optimization problem, i.e., as a search for a proper transformation that provides the best match between the images of the body and the scene. The paper describes an approach to human body registration that utilizes a hybrid evolutionary algorithm and image response analysis. Hybrid evolutionary algorithm provides an efficient procedure of global search in extremely large parameter space. Image response analysis allows to reduce the total amount of information that has to be processed during evaluation of potential solutions. In the process of the evolutionary search, response matrix of each template image is compared against response matrix of the reference image of the scene, in order to find the correct mapping between the images. The efficiency of the proposed approach is demonstrated on a test set of 2-D grayscale images.

Laser radar systems historically offer rich data sets for automatic target recognition (ATR). ATR algorithm development for laser radar has focused on achieving real-time performance with current hardware. Our work addresses the issue of understanding how much information can be obtain from the data, independent of any particular algorithm. We present Cramer-Rao lower bounds on target pose estimation based on a statistical model for laser radar data. Specifically, we employ a model based on the underlying physics of a coherent-detection laser radar. Most ATR algorithms for laser radar data are designed to be invariant with respect to position and orientation. Our information-theoretic perspective illustrates that even algorithms that do not explicitly involve the estimation of such nuisance parameters are still affected by them.