This paper describes a unified set of algorithms for the detection and discrimination of ground vehicles in second generation FLIR imagery. The computational substrate uses morphological signal processing and fractal analysis for a potential high-speed hardware implementation. The reported results are focused on the discrimination stage of the overall process. We demonstrate significant false alarm reduction through the use of textural and structural measures and a multistage discrimination procedure. The multistage discrimination process combines hard thresholding and fuzzy classification logic.

The subject of wavelet transform has received a lot of attention in research in the past few years. Wavelets are being widely applied in image compression, fractal analysis, speech synthesis, texture discrimination etc. In this paper, we discuss the feasibility of applying wavelet transformation to obtain a viable target feature vector for use in an automatic target recognition problem. The objective of target recognition problem is to discriminate the targets such as tanks, trucks, missile launchers, etc. from the background. We have applied our technique to the images supplied by the Air Force. The targets are segmented from the background clutter and subjected to wavelet decomposition. Energy values are computed for the wavelet decomposed target images and they form the target signature. Each target has a unique feature vector. Numerical results with field images are provided which show the potential of wavelets for feature extraction.

Automatic target detection and recognition (ATR) requires the ability to optimally extract the essential features of an object from (usually) cluttered environments. In this regard, efficient data representation domains are required in which the important target features are both compactly and clearly represented, enhancing ATR. Since both detection and identification are important, multidimensional data representations and analysis techniques, such as the continuous wavelet transform (CWT), are highly desirable. First we review some relevant properties of two 2D CWT. Then we propose a two-step algorithm based on the 2D CWT and discuss its adequacy for solving the ATR problem. Finally we apply the algorithm to various images.

The motion analysis of a time dependent signal in 1D of space and time using the spatio- temporal continuous wavelet transform [(1 + 1) CWT] allows signature detection of the moving objects. In this paper we first review some relevant properties of the (1 + 1)D CWT. We then apply it to motion analysis of some academic signals plunged in a cluttered medium. From there we infer a strategy for detection and recognition of moving targets. In the appendix we give analytical results of (1 + 1)D CWT applied to two basic signals: N ponctual moving objects, and N moving plane waves.

Automatic object recognition requirements exist to extract objects of interest from cluttered background. Most of the feature estimation and object extraction algorithms are designed to extract the known geometry of objects of interest. Recently, Markov random field modeling has been found useful in characterizing the clutter and the man-made objects. In this paper, a new approach to estimate features, particularly the first and second moments, is proposed to single out the most 'smooth' region in an image. The approach exploits the characteristics of man-made objects and background clutter characterized with natural scene using Markov random field modeling. Regions-of-interest are clustered and differentiated in terms of the features estimated at pixel level. Experiments have been conducted on different types of imagery such as camera, IR, and ladar. The results show that the algorithms work particularly well to extract man-made objects versus natural scene. Requiring no prior knowledge about the objects except that they can be characterized with relative 'smooth' surface, the algorithm is suitable for tracking an object moving in field. It is also useful for low/intermediate level processing for model-based pattern recognition systems.

Segmentation of targets embedded in clutter obtained by IR imaging sensors is one of the challenging problems in automatic target recognition (ATR). In this paper a new texture-based segmentation technique is presented that uses the statistics of 2D wavelet decomposition components of the lcoal sections of the image. A measure of statistical similarity is then used to segment the image and separate the target from the background. This technique is applied on a set of real sequential IR imagery and has shown to produce a high degree of segmentation accuracy across varying ranges.

Object recognition can be parametrized systematically through physically robust wave objects by linking features (observables) in scattered field data with features on the object (target) giving rise to the data. The wave objects are broadly separated into global (mode) and local (wavefront) categories. Their parametrization requires different wave-oriented signal- processing algorithms which are implemented conveniently in relevant subdomains of the configuration (space-time) spectrum (wavenumber-frequency) phase space. Projection of scattering data onto the phase space is achieved via Gaussian-windowed Fourier transforms, wavelet transforms, and windowed model-based (superresolution) algorithms. Example results are presented here for time-domain modes excited by an open cavity as well as by periodic and quasi-periodic structures, with data processed in the time-frequency phase space. Additionally, we consider frequency-domain modes (leaky modes supported by a dielectric slab) which are processed in the space-wavenumber phase space. For some situations, it is more appropriate to process the entire database simultaneously (without windowing), and we have used such techniques for certain modal and wavefront parametrizations. Concerning modal 'footprinting', results are presented for superresolution processing of measured short-pulse scattering data from resonant targets embedded in foliage (foliage penetrating radar); in these examples we extract late-time target resonant frequencies. We have also applied superresolution algorithms to wavefront-based processing, and results are presented here for model targets.

The recognition of 3D objects from images is an important problem in computer vision. Recently, the use of range images has gained increasing popularity because of the explicit representation of 3D shape information in range data. In our work we are using range images acquired by a sensor based on the coded light principle. This type of range scanner has recently become commercially available at low cost. But it suffers from the fact that no range data can be acquired in shadow areas of the input scene. Thus, the range data provided by the scanner is incomplete, in general. This fact leads to significant complications in object recognition. In the present paper we describe a model-based recognition system that is robust if there are regions of missing data in an input range image. The system uses a graph matching technique that searches for an optimal correspondence between surface patches extracted from an input range image and objects from a model library. In the model matching process, a hierarchy of constraints aiming at the recognition of objects with few distortions first. As long as there are unmatched surface patches in the input data, we gradually relax the constraints. The system works for both objects with planar and curved surfaces. In practical experiments, it has shown a high degree of robustness on images of complex scenes in which significant portions of the data were missing. A correct recognition rate of 92% with 8% of the objects being rejected has been achieved on a test set.

A variety of experimental results indicate that dolphins possess a unique and sophisticated sonar system. In addition, this sonar system is highly adaptive in detecting, discriminating, and recoginizing objects in highly reverberating and noisy environments. In this paper a new approach using resonance scattering theory in target detection and recognition is presented. The results seems to imply that this approach may be useful in minelike target detection and identification.

The new normal-mode method SWAMP (shallow water acoustical mode propagation) is used to investigate the high-frequency sound propagation in oceanic waveguides with pronounced ducts. Both the monochromatic and pulse signals are considered. It is shown that the time domain maps representing the arrived signal intensity relative to time and depth contain information to reconstruct the vertical sound velocity profile, source location, and bottom properties.

In this paper a neural-network-based automatic target recognition (ATR) classifier is developed. The ATR classifier consists of a learning vector quantization (LVQ) algorithm followed by a multilayer perceptron (MLP). The LVQ is used as the feature extractor and the MLP as the classifier. The LVQ algorithm adaptively extracts a set of target templates (centroids) that are assumed to represent the target signatures. The Euclidean distances between the centroids and the input target are passed to an MLP. The MLP uses these distances as input and performs a classification. Experimental results are presented for two different test sets. The first test set has similar characteristics to those of the training set, and the ATR classifier does very well. However, the second test set has a different characteristics and the ATR classifier performance is poor.

Face recognition has been a challenging task in academic research and in industrial applications. Defined by many distinct features (and the relative spatial position of those features), faces are very complex targets. Small inter-facial differences with high intra-facial variability makes face recognition a particularly complex task. Until now, face recognition systems have relied on input from the visual domain. The neural network system reported in this paper recognizes faces based on sonar input. Research on echolocating bats has demonstrated that bats can use sonar to accurately perceive detailed descriptions of objects. Previous research has shown that a sonar neural network system can recognize simple, 3D targets regardless of orientation (Dror, et al., 1995). In the present study we examine the effectiveness of using sonar input for more complex target recongition tasks. We use sonar echoes from faces (recorded in a variety of facial expressions) to train a neural network to recognize faces, regardless of facial expressions. After training, we examine the network's ability to generalize and correctly recognize the faces based on echoes from novel facial expressions, which were not included in the training set. The performance of the network on these novel echoes was 100% correct. To insure that our results were not due to the specific faces we used, we replicated our results two more times using different faces. Again, performance was almost perfect--99.6% and 100%. The results show that a neural network can learn to recognize faces based on sonar, and demonstrate that sonar can be a very effective input for neural networks that perform pattern recognition tasks.

Conventional target tracking algorithms based on linear estimation techniques perform quite efficiently when the target motion does not involve maneuvers. Target maneuvers involving short term accelerations, however, cause a bias (e.g. jump) in the measurement sequence, which unless compensated, results in divergence of the Kalman filter that provides estimates of target position and velocity, in turn leading to a loss of track. Accurate compensation for the bias requires processing more samples of the input signals which adds to the computational complexity. The waiting time for more samples can also result in a total loss of track since the target can begin a new maneuver and if the target begins a new maneuver before the first one is compensated for, the filter would never converge. Most of the proposed algorithms in the current literature hence have the disadvantage of losing the target in short term accelerations, i.e., when the duration of acceleration is comparable to the time period between the measurements. The time lag for maneuver modelings, which have been based on Bayesian probability calculations and linear estimation shall propose a neural network scheme for the modeling of target maneuvers. The primary motivation for employing compensation. The parallel processing capability of a properly trained neural network can permit fast processing of features to yield correct acceleration estimates and hence can take the burden off the primary Kalman filter which still provides the target position and velocity estimates.

Many military systems have a critical need for aided target recognition, or cuing. This includes several systems with wide field-of-view search missions such as the UAV, EFOG-M, and Comanche. This report discusses one new approach: a multiple region of interest processor based on diffraction pattern sampling and digital neural network processing. In this concept an optical system segments the image into multiple, rectangular regions of interest and in parallel converts each ROI, be it visible, IR, or radar, to a spatial frequency power spectrum and samples that spectrum for 64 features. A neural network learns to correlate those features with target classes or identifications. A digital system uses the network weights to recognize unknown targets. The research discussed in this report using a single ROI processor showed a very high level of performance. Out of 1024 trials with models of five targets of F- 14, F-18, HIND, SCUD, and M1 tanks, there were 1023 correct classifications and 1 incorrect classification. Out of 1514 trials with those images plus 490 real clutter scenes, there were 1514 correct decisions between target or no-target. Of the 1024 target detections, there were 1023 correct classifications. Out of 60 trials with low resolution IR images of real scenes, there were 60 correct decisions between target and no-target. Of the 40 target detections, there were 40 correct classifications.

The motive for this study is to achieve near aspect-independent target identification with the assumption that radar signatures have fuzzy memberships in more than one class. Radar cross section measurements of four commercial aircraft, obtained experimentally in a compact range, are used for training and testing a three layered back propagation neural network for target identification purposes. It is assumed that the aspect angles (or azimuth positions) of all four targets are known within 20 degrees uncertainty range. The radar cross section parameters of each target are assigned to three fuzzy membership functions, and for each combination of membership functions there is a desired fuzzy output. The performance of the proposed target recognition system is examined assuming different noise scenarios and various levels of azimuth ambiguity. The proposed fuzzy neural scheme is also tested in scenarios where the maximum likelihood is available and the performances of both recognition techniques are compared. It is assumed that radar target signatures have fuzzy class separation. Issues concerning the number of hidden nodes, training parameters, and weight convergence are discussed.

One approach to making a transient radiating and/or receiving antenna, involves a timed array which allows one to steer the beam without physically rotating the antenna. An array of interconnected elements (unit cells) can be used to give a very broadband characteristic. Various designs of such unit cells are reviewed and discussed for single and dual polarization. The low-frequency part of the pulse spectrum is governed by the electric and magnetic dipole moments dipole moments of the array and additional conductors (impedance loaded) connected to the array.

We study the scattering interaction of electromagnetic pulses of short duraction with two targets of simple shape. The targets are metal spheres buried at different depths in sand contained in an indoor sandbox of length and width of about 5 m and depth of 2 m. The spectrum of the backscattered echo is determined from measured data when the target is illuminated by an impulse radar system. In general, however, the signature features extracted when the antenna and the target are located in different half-spaces having different electrical properties are distorted by multiple scattering becomes less significant and, as the distance increases, the backscattering signature will ultimately approach that of the target in free space. We examine here the effect of multiple scattering on the target's signature will ultimately approach that of the target in free space. We examine here the effect of multiple scattering on the target's signatures when they are buried at different depths. In particular, we examine how the signature features evolve in time, both the primary signature feature and the subsequent features caused by multiple scattering. This is carried out using a pseudo-Wigner distribution (PWD) with a Gaussian time-window, having its width properly adjusted to suppress the interference of cross-terms in the PWD, yet retaining the desired property of concentrating the extracted features in the time-frequency domain. The results serve to assess the possibility if identifying subsurface targets using a ground penetrating radar (GPR).

We review techniques to analyze the backscattering cross-sections of plane electromagnetic waves of wavelength (lambda) returned from cylinders of finite length. The various appropriate techniques yield considerably different results depending on the values of two physical ratios, 2b/(lambda) and a/(lambda) , where 2b is the cylinder's length and 'a' is its radius. The relevant regions are those for which 2b/(lambda) is either much larger than, near to, or much smaller than unity, while at the same time the a/(lambda) ratio is either much smaller than, near to, or much larger than unity. There are thus, nine such regions. Closed- form analytical solutions are not possible in some of those regions, and in many instances the expressions for the cross-section contain unexpectedly cumbersome complexities. These cases are summarized and diplayed in a 2b/(lambda) vs. a/(lambda) plane, in which we indicate the regions where polarization effects are, or are not, significant. The basic technique in the Mie (or resonance) region, a/(lambda) approximately 1, is the exact solution of the governing integral equations. In the Rayleigh (i.e., ka << 1), or in the geometrical optics (i.e., ka $GT$GT 1) regions, we show appropriate low or high-frequency asymptotic approximations. A particular goal of this work is to determine the scattering features or characteristics of thin long rods, normal to perfectly conducting planes, that could model periscopes protruding out of an ideally flat ocean, with a view toward the remote and unambiguous characterization of such objects. The resulting formulas are plotted in various pertinent situations.

A practical method for the recognition of armored vehicles in FLIR (forward-looking infrared radar) imagery is presented. This unique method is applied to second-generation IR images of unoccluded vehicles. The principal value of our technique is that recognition is invariant to large changes (up to 45 degrees) in target depth aspect (rotation into or out of the view plane), as well as large changes in range from sensor to target. Feature detection is robust, and is not easily fooled by various irregularities, appendages, and variations among the test targets. This is accomplished by ensuring that certain geometric relationship constraints are satisfied. Various length measurements from target centroid and perimeter are combined into six ratios to supply input features. Such simple, reliable measurements provide low computational complexity, thus having the potential to scale well to larger target sets. Real-time implementation may also be possible. The principal limitation is that, thus far, we have tested the system on less than a dozen different vehicle types. Nevertheless, discrimination between similar-looking vehicles was very high. Also the algorithm works on armored vehicles tha have markedly different appearances (e.g., with or without turrets). Currently our method is reliable with up to 10% shot-noise appearing in the scene. Pattern classification is achieved using an expert system, but this has not been completely automated. We are now using the CLIPS expert system shell, and heuristically coding the various allowable feature ranges. EMYCIN type confidence factors are also being used to give a quantitative measure of classification.

The purpose of this paper is to address the use of an autoregressive method to discern scattering center distribution. This information serves the purpose of a filter that is utilized to determine the minimum-weighted-norm least-squares solution of the autoregressive parameters. The parameters are employed to extrapolate a set of wideband target signatures collected in the spatial-frequency domain. The periodogram of these signatures yields a profile history or an image with high resolution and accurate reflectivity measurements. The iterative autoregressive method is compared to the principal component modified covariance to illustrate performance improvement.

The US Army is currently investigating the use of acoustics to locate and discriminate tank rounds fired during live-fire training exercises for armored vehicle gunners. Being able to determine round type and location at the intended target greatly enhances training realism. By determining a round's velocity, miss distance from a sensor and acoustic shock wave duration, the round can be classified into one of five size categories. The round's location at the target is calculated as a by-product during the determination of miss distance. A prototype acoustic discrimination system has been developed that consists of a microprocessor with multichannel A/D hardware and a few acoustic sensors. For each round fired, the prototype system calculates a projectile shape value (PSV) from the round's velocity, miss distance, and shock wave duration. The PSV is then compared to a look-up table to classify the round according to size. Initial tests of the prototype acoustic round discrimination system have been successful.

The system created has an accuracy of 0.05% for single measurements. It is reached due to controlled electro-optical polarization probing and recieving channels of the system. This makes it possible to fully automate processes of measurement and calibration of system. It also makes available operational elasticity of the system: choice of measurement parameters (whole Stokes vector, Mueller matrix, or particularly some of their elements) only by means of software. Measurement time reached a few microseconds. Exceptions of influence of optical radiation background component of measurement results without use of interference filter is a distinction of this system. It considerably increases the level of interference suppression. It is possible due to intensity modulation of optical radiation is sysem probing channel with simultaneous modulation of polarization. Other distictions of this system include the concept of processing received information. On the basis of this concept the following statement is laid: an arbitrary Mueller matrix is always presented on basis of amplitude and phase anisotropy matrices. There are some very important consequences of this statement. Namely it allows us to characterize suitably the physical behavior of the object with respect to polarization phenomena and to perform a general classification of objects based on the above mentioned approach. Furthermore, it is possible on this basis to synthesize the polarization systems with given characteristics.

This paper illustrates the performance achieved using a new multiclass distance classifier correlation filter (DCCF) design. Previous distance classifier methodology' was directly applicable to only two classes at a time. This paper shows how this new multiclass method distinguishes several classes with actual imaging data sequences until spatial resolution limits are reached. Keywords : synthetic discriminant function ( SDF) , distance classifier correlation filter (DCCF) , multiclass object recognition, multiclass ATR

Because of the availability of explicit 3D information, the use of range data has become quite popular in computer vision. One of the common methods for range data acquisition is based on coded light. An advantage of this method is that it yields not only a range image of a scene, but also a greylevel image that is in registration with the range image. Thus, information from the greylevel and range image can be easily combined. On the other hand, the coded light approach suffers from the existence of shadow areas, where no range data can be measured. That is, it gives only an incomplete range image, in general. In this paper, we present a new approach to the interpretation of edges in range images of polyhedra. First, we integrate the edges extracted from the greylevel and range image of a scene. Then, the edges are classified into one of the types jump, convex, concave, or non-geometric. While jump, convex, and concave edges correspond to real edges, the non-geometric edges, which are caused by shadow, can be removed. Such a classification of physical edges together with the elimination of shadow edges potentially improves any subsequent object recognition step.

Power law features obtained from fractal dimension analysis have proved to be useful for the discrimination of textures. We consider texture to be a distributed, regional visual effect--an identically distributed random field. Contamination and misclassification occur as the calculations for fractal dimension encompass an area of the image that is not texturally homogeneous. The random field in that region is not identically distributed. Boundary information can be utilized to preserve homogeneity by adapting the subfield being processed. This will improve the texture information near the boundary and can vastly improve the accuracy of the segmentation of the texture regions for scene analysis.

In this paper, the use of ESPRIT is explored with relatively small baseline acoustic sensor arrays for passive direction-of-arrival (DOA) estimation and tracking of ground vehicles in the battlefield environment. Classical delay-sum beamformers do not provide high-resolution DOA estimates and their ability to resolve closely-spaced targets is limited. This deficiency is due to the geometric constraint of the sensor array; the baseline cannot be expanded beyond a few meters due to the lack of spatial coherence. ESPRIT is attractive for acoustic array processing, operating at a lower computational rate and less sensitive to sensor array imperfections than comparable high-resolution DOA algorithms. Preliminary experimental results obtained from the application of narrowband ESPRIT for a single tracked vehicle, which is a wideband acoustic source, look promising.

This paper addresses object tracking problems in an image sequence using an active contour model called '(rho) -snake'. This model is the result of the combination of classical snakes and elements from the robust estimators theory. Snakes are energy-minimizing curves with global constraints that segment deforming shapes. The theory of robust estimators provides a framework that makes parameter estimation free from outliers. We have introduced (rho) - snakes to use these two techniques to achieve a goal: tracking a moving shape along an image sequence without being influenced by erroneous information of images. Attempting to imporve this new technique, we present parallel processing and a faster way of implementing (rho) - snakes. We also have defined robust energies, both spatial and temporal. As these energies include prediction, they fit our problem: tracking poor contrasted and fast moving object in a noisy IR sequence.

This paper describes the US Army's efforts under the ATR relational template matching program to develop a multi-sensor forward-looking IR (FLIR) and millimter wave (MMW) radar automatic target recognition (ATR) algorithm. The general problem consists of identifying ground targets at low depression angles of less than 6 degrees and ranges of 500 to 6000 m. However, the algorithm is directly applicable to other target sets, including air targets, and ground targets at high depression angles. Our phase I goal was to provide a proof- of-principle demonstration of this new model-based methodology. We did this on simulated FLIR imagery for a three target class problem. A comprehensive test and evaluation was performed at the Night Vision and Electronic Sensors Directorate. Our goal in phase II was to demonstrate this new algorithm approach on realistic field-collected second-generation FLIR imagery with a larger target set consisting of ten ground vehicles. With phase III, our goal was to demonstrate multisensor fusion on FLIR (first generation) fused with laser radar. This was demonstrated on a four-ground-target-class problem. The limitations to first-generation FLIR data and four targets was due purely to restrictions of the field-collected data. Our goal in phase IV is to demonstrate this methodology for FLIR fused with MMW radar data. Phase IV will be a three-year effort, with the first year concentrating on using the MMW radar data for detection, reduction in false alarms, and the provision of accurate range (i.e., scale) to the FLIR algorithm. The goal of the second and third years will be to fuse the MMW data into not only the detection phase of the algorithm but also the recognition and identification portions of the algorithm. This paper presents the general model-based methodology used for the FLIR and the neural network methodology used to provide the FLIR with the MMW detections and scale. Results are provided for the FLIR-only algorithm for short range and moderate range scenarios given an accurate estimate of target range.

This paper compares the performance of multitarget track initiation techniques in clutter environments using real radar data collected by an operating radar system. The track initiation problem is considered a problem of detection, and the performance is evaluated using the receiver operating characteristics. Four track initiation techniques, namely the heuristic rule method, the logic-based method, the Hough transform approach and the modified Hough transform method, are considered. The first two methods are sequential and the last two are batch technique, which are the two major categories of track initiation methods. The effect of the number of radar scans used to initiate a track, the effect of the initiation method to a tracker, and the comparison of the four track initiation methods are analyzed to understand the practicalities of these track initiation techniques.

Our pattern theoretic approach to the automated understanding of forward-looking infrared (FLIR) images brings the traditionally separate endeavors of detection, tracking, and recognition together into a unified jump-diffusion process. New objects are detected and object types are recognized through discrete jump moves. Between jumps, the location and orientation of objects are estimated via continuous diffusions. An hypothesized scene, simulated from the emissive characteristics of the hypothesized scene elements, is compared with the collected data by a likelihood function based on sensor statistics. This likelihood is combined with a prior distribution defined over the set of possible scenes to form a posterior distribution. The jump-diffusion process empirically generates the posterior distribution. Both the diffusion and jump operations involve the simulation of a scene produced by a hypothesized configuration. Scene simulation is most effectively accomplished by pipelined rendering engines such as silicon graphics. We demonstrate the execution of our algorithm on a silicon graphics onyx/reality engine.

The problem of resoration of inhomogeneity spatial distribution in the ocean is investigated using methods of hydroacoustical vision. In all schemes the dark field method is applicated. Distributions obtained as a results of coherent and incoherent summation of partial images are compared.

The dynamic acoustic information processed by a sonar array is of critical importance to target discrimination and recognition. A visualization of the true near-field acoustic energy distribution of an active sonar array would be of great value to understanding and characterizing the effects of structural defects in the array and its interface to the water medium. Knowledge of the projected near-field acoustic energy would also enhance the analysis of returned sonar signals for homing and target recognition. Holographic interferometry has presented itself as a viable and useful method for the realization of this type of information.

Mathematical morphology based on set theory has been applied to various areas in image processing. Pita proposed an object recognition algorithm using morphological shape representation (MSR). The MSR is a representation scheme to combine constructive solid geometry and MSD. Pita's algorithm is a simple and adequate approach to recognize objects that are rotated 45 degree-units with respect to the model object. However, this recognition scheme fails in the case of random rotation. This disadvantage may be compensated by defining small angle increments. However, this solution may greatly increase computational complexity because the smaller the step, more number of rotations needed. In this paper, we propose a new method for object recognition based on MSD. The first step of our method decomposes a binary shape into a union of simple binary shapes, and then a new tree structure is constructed which can represent the relations of binary shapes in an object. Finally, we obtain the feature information invariant to the rotation, translation, and scaling from the tree, and calculate matching scores using efficient matching measure. Because our method does not need to rotate the object to be tested, it could be more efficient than Pita's method. MSR has an intricate structure so that it might be difficult to calculate matching scores even for a little complex object. But our tree has a simpler structure than MSR, and it is easier to calculate the matching score. We experimented on 20 test images, and scaled, rotated, and translated versions of five kinds of automobile images. The simulation result using octagonal structure elements shows 95% recognition rate. The experimental results using approximated circular structure elements are examined. ALso, the effect of noise on MSR scheme is discussed.

The reliable long-range detecting and tracking of fast moving targets is a crucial problem in shooting range tests. The target signal received by IR tracking equipment usually changes instantaneously in both time domain and in the space domain, and it will by buried in the background noise more seriously when the projected target moves farther. A novel method of target signal detection is proposed by using nonlinear feather accumulation, 2D signal-feature joint detection, temporal association, and synthetical competition decision techniques. The developed algorithm can not only overcome the detection missing problem resulted from the spatial and temporal violent undulating of target signal, but also guarantee along-range optimal detection results. It has been used to increase the target detection range greatly and can acquire the weak target signal effectively even under about -3dB signal to noise ration (SNR).

The performance of automatic target recognition (ATR) systems using thermal infrared images is limited by the low contrast in intensity for terrestrial scenes. We are developing a thermal imaging technique where, in each image pixel, a combination of intensity and polarization data is captured simultaneously. In this paper, we demonstrate, using synthetic polarization images, that a combination of intensity and polarization data will significantly improve the performance of detection and classification functions in an ATR system. The images were generated using a ray tracing computer program, modified to calculate the polarization characteristics of thermal radiation emitted from surfaces. We developed novel polarization- sensitive target edge detection, segmentation, and recognition algorithms. A set of performance metrics for the evaluation showed that, for large ranges of viewing elevation and aspect angles, using a combination of polarization and intensity data significantly improves the performance of the algorithms over using only the intensity data.