This PDF file contains the front matter associated with SPIE Proceedings Volume 6967, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and the Conference Committee listing.

For over half a century, scientists and engineers have worked diligently to advance computational intelligence. One application of interest is how computational intelligence can bring value to our war fighters. Automatic Target Recognition (ATR) and sensor fusion efforts have fallen far short of the desired capabilities. In this article we review the capabilities requested by war fighters. When compared to our current capabilities, it is easy to conclude current Combat Identification (CID) as a Family of Systems (FoS) does a lousy job. The war fighter needed capable, operationalized ATR and sensor fusion systems ten years ago but it did not happen. The article reviews the war fighter needs and the current state of the art. The article then concludes by looking forward to where we are headed to provide the capabilities required.

Automatic target detection (ATD) systems process imagery to detect and locate targets in support of intelligence, surveillance, reconnaissance, and strike missions. Accurate prediction of ATD performance would assist in system design and trade studies, collection management, and mission planning. Specifically, a need exists for ATD performance prediction based exclusively on information available from the imagery and its associated metadata. In response to this need, we undertake a modeling effort that 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 ATD algorithms, which is even applicable when no image truth is available (e.g., when evaluating denied area imagery). Ongoing efforts to develop such a prediction model have met with some success. Previous results presented models to predict performance for several ATD methods. This paper extends the work in several ways: extension to a new ATD method, application of the modeling to a new image set, and an investigation of systematic changes in the image properties (resolution, noise, contrast). The paper concludes with a discussion of future research.

We consider automatic target recognition (ATR) in infrared (IR) imagery using the minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF). As in our prior work (SPIE 6566-03), we consider classification of true-class CAD targets and rejection of real clutter and unseen confuser CAD objects with range and full 360° aspect view variations. In this work, we address rejection of new UCIR bush clutter data. We also present performance scores for several different training and test cases with attention to filter capacity, i.e., the number of training images that can be included in one filter before performance on the test set deteriorates appreciably. We find that range rather than aspect view distortions seem to affect filter capacity more. Initial target contrast ratio tests are also presented. To more properly address clutter, in all tests we now form the magnitude of the output correlation plane before analysis. We also address when and why linear versus circular correlations are best. We also address DIF filter-synthesis and fast implementation for wide area "search" test regions. This introduces new issues concerning the region over which correlation plane energy is minimized in filter synthesis and the size of the FFT to use in tests. A key issue is that both training and tests should use the same procedures. This is vital for training and test metrics to be comparable. We distinguish between whether linear or circular correlation plane energy is minimized.

A generic nonlinear dynamic range compression deconvolver (DRCD) is proposed. We have performed the dynamic range compression deconvolution using three forms of nonlinearities: (a) digital implementation- A-law/μ-law, (b) hybrid digital-optical implementation- two-beam coupling photorefractive holography, and (c) all optical implementation- MEMS deformable mirrors. The performance of image restoration improves as the saturation nonlinearity increases. The DRCD could be used as a preprocessor for enhancing Automatic Target Recognition (ATR) system performance. In imaging through atmosphere, factors such as rain, snow, haze, pollution, etc. affect the received information from a target; therefore the need for correcting these captured images before an ATR system is required. The DRCD outperforms well-established image restoration filters such as the inverse and the Wiener filters.

The last five years have seen a renewal of Automatic Target Recognition applications, mainly because of the latest advances in machine learning techniques. In this context, large collections of image datasets are essential for training algorithms as well as for their evaluation. Indeed, the recent proliferation of recognition algorithms, generally applied to slightly different problems, make their comparisons through clean evaluation campaigns necessary. The ROBIN project tries to fulfil these two needs by putting unclassified datasets, ground truths, competitions and metrics for the evaluation of ATR algorithms at the disposition of the scientific community. The scope of this project includes single and multi-class generic target detection and generic target recognition, in military and security contexts. From our knowledge, it is the first time that a database of this importance (several hundred thousands of visible and infrared hand annotated images) has been publicly released. Funded by the French Ministry of Defence (DGA) and by the French Ministry of Research, ROBIN is one of the ten Techno-vision projects. Techno-vision is a large and ambitious government initiative for building evaluation means for computer vision technologies, for various application contexts. ROBIN's consortium includes major companies and research centres involved in Computer Vision R&D in the field of defence: Bertin Technologies, CNES, ECA, DGA, EADS, INRIA, ONERA, MBDA, SAGEM, THALES. This paper, which first gives an overview of the whole project, is focused on one of ROBIN's key competitions, the SAGEM Defence Security database. This dataset contains more than eight hundred ground and aerial infrared images of six different vehicles in cluttered scenes including distracters. Two different sets of data are available for each target. The first set includes different views of each vehicle at close range in a "simple" background, and can be used to train algorithms. The second set contains many views of the same vehicle in different contexts and situations simulating operational scenarios.

Over the five past years, the computer vision community has explored many different avenues of research for Automatic Target Recognition. Noticeable advances have been made and we are now in the situation where large-scale evaluations of ATR technologies have to be carried out, to determine what the limitations of the recently proposed methods are and to determine the best directions for future works. ROBIN, which is a project funded by the French Ministry of Defence and by the French Ministry of Research, has the ambition of being a new reference for benchmarking ATR algorithms in operational contexts. This project, headed by major companies and research centers involved in Computer Vision R&D in the field of Defense (Bertin Technologies, CNES, ECA, DGA, EADS, INRIA, ONERA, MBDA, SAGEM, THALES) recently released a large dataset of several thousands of hand-annotated infrared and RGB images of different targets in different situations. Setting up an evaluation campaign requires us to define, accurately and carefully, sets of data (both for training ATR algorithms and for their evaluation), tasks to be evaluated, and finally protocols and metrics for the evaluation. ROBIN offers interesting contributions to each one of these three points. This paper first describes, justifies and defines the set of functions used in the ROBIN competitions and relevant for evaluating ATR algorithms (Detection, Localization, Recognition and Identification). It also defines the metrics and the protocol used for evaluating these functions. In the second part of the paper, the results obtained by several state-of-the-art algorithms on the SAGEM DS database (a subpart of ROBIN) are presented and discussed

Existing techniques for hyperspectral image (HSI) anomaly detection are computationally intensive precluding real-time implementation. The high dimensionality of the spatial/spectral hypercube with associated correlations between spectral bands present significant impediments to real time full hypercube processing that accurately encapsulates the underlying modeling. Traditional techniques have imposed Gaussian models, but these have suffered from significant computational requirements to compute large inverse covariance matrices as well as modeling inaccuracies. We have developed a novel data-driven, non-parametric HSI anomaly detector that has significantly reduced computational complexity with enhanced HSI modeling, providing the capability for real time performance with detection rates that match or surpass existing approaches. This detector, based on the Support Vector Data Description (SVDD), provides accurate, automated modeling of multi-modal data, facilitating effective application of a global background estimation technique which provides the capability for real time operation on a standard PC platform. We have demonstrated one second processing time on hypercubes of dimension 256×256×145, along with superior detection performance relative to alternate detectors. Computation performance analysis has been quantified via processing runtimes on a PC platform, and detection/false-alarm performance is described via Region Operating Characteristic (ROC) curve analysis for the SVDD anomaly detector vs. alternate anomaly detectors.

Recently, the 1-D spectral fringe-adjusted joint transform correlation (SFJTC) technique has been proposed as an effective means for performing deterministic target detection in hyperspectral imagery. In this work, we explore the use of the discrete wavelet transform (DWT) as a pre-processing tool for SFJTC-based target detection. To quantify improvement and compare performance in the detection process, receiver operating characteristic (ROC) curves are generated and the areas under the ROC curves (AUROC) are computed. The basic premise of this work is that selected coefficients generated from a desired level of the DWT decomposition of the data can be used in place of the original data for improved SFJTC-based detection. We illustrate this by conducting experiments on two different hyperspectral scenes containing varying amounts of simulated noise. Results indicate that use of the DWT coefficients significantly improves the detection performance, especially in the presence of noise.

A novel variational method using level sets that incorporate spectral angle distance in the model for automatic target detection is presented. Algorithms are presented for detecting both spatial and pixel targets. The new method is tested in tasks of unsupervised target detection in hyperspectral images with more than 100 bands, and the results are compared with a widely used region-based level sets algorithm. Additionally, techniques of band subset selection are evaluated for the reduction of data dimensionality. The proposed method is adapted for supervised target detection and its performance is compared with traditional orthogonal subspace projection and constrained signal detector for the detection of pixel targets. The method is evaluated with different complexity such as noise levels and target sizes.

Hyperspectral data (HS) is increasingly used in target detection applications since it provides both spatial and spectral information about the scene. One of the main challenges in HS data is to handle a large volume of data. On the other hand, mutispectral data provides the information with reduced number of bands. As a result, target detection in multispectral image is more challenging due to lack of information about the objects. In this paper, we presented a new approach to detect land mines in multispectral images. We showed that application of matched filter (MF) to multispectral data is not suitable to detect the targets but after selecting some features based on principal component analysis (PCA) enables it to detect all the targets. We also described a segmentation technique-sliding concentric window (SCW) to extract the land mines from the clutter.

Three dimensional imaging is a powerful tool for object detection, identification, and classification. 3D imaging allows removal of partial obscurations in front of the imaged object. Traditional 3D image sensing has been Laser Radar (LADAR) based. Active imaging has benefits; however, its disadvantages are costs, detector array complexity, power, weight, and size. In this keynote address paper, we present an overview of 3D sensing approaches based on passive sensing using commercially available detector technology. 3D passive sensing will provide many benefits, including advantages at shorter ranges. For small, inexpensive UAVs, it is likely that 3D passive imaging will be preferable to active 3D imaging.

Feature-aided target verification is a challenging field of research, with the potential to yield significant increases in the confidence of re-established target tracks after kinematic confusion events. Using appropriate control algorithms airborne multi-mode radars can acquire a library of HRR (High Range Resolution) profiles for targets as they are tracked. When a kinematic confusion event occurs, such as a vehicle dropping below MDV (Minimum Detectable Velocity) for some period of time, or two target tracks crossing, it is necessary to utilize feature-aided tracking methods to correctly associate post-confusion tracks with pre-confusion tracks. Many current HRR profile target recognition methods focus on statistical characteristics of either individual profiles or sets of profiles taken over limited viewing angles. These methods have not proven to be very effective when the pre- and post- confusion libraries do not overlap in azimuth angle. To address this issue we propose a new approach to target recognition from HRR profiles. We present an algorithm that generates 2-D imagery of targets from the pre- and post-confusion libraries. These images are subsequently used as the input to a target recognition/classifier process. Since, center-aligned HRR Profiles, while ideal for processing, are not easily computed in field systems, as they require the airborne platform's center of rotation to line up with the geometric center of the moving target (this is impossible when multiple targets are being tracked), our algorithm is designed to work with HRR profiles that are aligned to the leading edge (the first detection above a threshold, commonly referred to as Edge-Aligned HRR profiles). Our simulated results demonstrate the effectiveness of this method for classifying target vehicles based on simulations using both overlapping and non-overlapping HRR profile sets. The algorithm was tested on several test cases using an input set of .28 m resolution XPATCH generated HRR profiles of 20 test vehicles (civilian and military) at various elevation angles.

This paper primarily investigates the use of shape-based features by an Automatic Target Recognition (ATR) system to classify various types of targets in Synthetic Aperture Radar (SAR) images. In specific, shapes of target outlines are represented via Elliptical Fourier Descriptors (EFDs), which, in turn, are utilized as recognition features. According to the proposed ATR approach, a segmentation stage first isolates the target region from shadow and ground clutter via a sequence of fast thresholding and morphological operations. Next, a number of EFDs are computed that can sufficiently describe the salient characteristics of the target outline. Finally, a classification stage based on an ensemble of Support Vector Machines identifies the target with the appropriate class label. In order to experimentally illustrate the merit of the proposed approach, SAR intensity images from the well-known Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset were used as 10-class and 3-class recognition problems. Furthermore, comparisons were drawn in terms of classification performance and computational complexity to other successful methods discussed in the literature, such as template matching methods. The obtained results portray that only a very limited amount of EFDs are required to achieve recognition rates that are competitive to well-established approaches.

Template-based classification algorithms used with synthetic aperture radar (SAR) automatic target recognition (ATR) degrade in performance when used with spatially mismatched imagery. The degradation, caused by a spatial mismatch between the template and image, is analyzed to show acceptable tolerances for SAR systems. The mismatch between the image and template is achieved by resampling the test imagery to different pixel spacings. A consistent SAR dataset is used to examine pixel spacings between 0.1069 and 0.2539 meters with a nominal spacing of 0.2021 meters. Performance degradation is observed as the pixel spacing is adjusted, Small amounts of variation in the pixel spacing cause little change in performance and allow design engineers to set reliable tolerances. Alternatively, the results show that using templates and images collected from slightly different sensor platforms is a very real possibility with the ability to predict the classification performance.

This paper presents a preliminary study of information-theoretic divergence between sets of LADAR image data. This study has been motivated by the hypothesis that despite the huge dimensionality of raw image space, related images actually lie on embedded manifolds within this set of all possible images and can be represented in much lowerdimensional sub-spaces. If these low-dimensional representations can be found, information theoretic properties of the images can be exploited while circumventing many of the problems associated with the so-called "curse of dimensionality." In this study, PCA techniques are used to find a low-dimensional sub-space representation of LADAR image sets. A real LADAR image data set was collected using the AFSTAR sensor and a synthetic image data set was created using the Irma LADAR image modeling program. One unique aspect of this study is the use of an entirely synthetic data set to find a sub-space representation that is reasonably valid for both the synthetic data set and the real data set. After the sub-space representation is found, an information-theoretic density divergence measure (Cauchy- Schwarz divergence) is computed using Parzen window estimation methods to find the divergence between and among the sets of synthetic and real target classes. These divergence measures can then be used to make target classification decisions for sets of images. In practice, this technique could be used to make classification decisions on multiple images collected from a moving sensor platform or from a geographically distributed set of cooperating sensor platforms operating in a target region.

This paper proposes a radar target identification system using down range profile radar signatures. The recognition is performed using a multi-layer perceptron trained via particle swarm optimization (PSO). The recognition results are compared with those obtained using back propagation training. The paper also uses PSO for modeling target signatures and extracting target scattering centers assuming that they can be modeled as an auto regressive moving average model. Real radar signatures of commercial aircraft are used to assess the performance of the techniques proposed. The results focus on comparing PSO based techniques with others used for target modeling and recognition.

Underwater mine identification persists as a critical technology pursued aggressively by the Navy for fleet protection. As such, new and improved techniques must continue to be developed in order to provide measurable increases in mine identification performance and noticeable reductions in false alarm rates. In this paper we show how recent advances in the Volume Correlation Filter (VCF) developed for ground based LIDAR systems can be adapted to identify targets in underwater LIDAR imagery. Current automated target recognition (ATR) algorithms for underwater mine identification employ spatial based three-dimensional (3D) shape fitting of models to LIDAR data to identify common mine shapes consisting of the box, cylinder, hemisphere, truncated cone, wedge, and annulus. VCFs provide a promising alternative to these spatial techniques by correlating 3D models against the 3D rendered LIDAR data.

Using matched filters to find targets in cluttered images is an old idea. Human operators can interactively find threshold values to be applied to the correlation surface that will do a good job of binarizing it into signal/non-signal pixel regions. Automating the thresholding process with nine measured image statistics is the goal of this paper. The nine values are the mean, maximum, and standard deviation of three images: the input image presumed to have some signal, an NxN matched filter kernel in the shape of the signal, and the correlation surface generated by convolving the input image with the matched filter kernel. Several thousand input images with known target locations and reference images were run through a correlator with kernels that resembled the targets. The nine numbers referred to above were calculated in addition to a threshold found with a time consuming brutal algorithm. Multidimensional radial basis functions were associated with each nine number set. The bump height corresponded to the threshold value. The bump location was within a nine dimensional hypercube corresponding to the nine numbers scaled so that all the data fell within the interval 0 to 1 on each axis. The sigma (sharpness of the radial basis function) was calculated as a fraction of the squared distance to the closest neighboring bump. A new threshold is calculated as a weighted sum of all the Gaussian bumps in the vicinity of the input 9D vector. The paper will conclude with a table of results using this method compared to other methods.

Many traditional methods produce classification results by processing one image frame at a time. For instance, conventional correlation filters are designed to yield well defined correlation peaks when a pattern or object of interest is present in the input image. However, the decision process is memory-less, and does not take advantage of the history of results on previous frames in a sequence. Recently, Kerekes and Kumar introduced a new Bayesian approach for multi-frame correlation that first produces an estimate of the object's location based on previous results, and then builds up the hypothesis using both the current data as well as the historical estimate. A motion model is used as part of this estimation process to predict the probability of the object at a particular location. Since the movement and behavior of objects can change with time, it may be disadvantageous to use a fixed motion model. In this paper, we show that it is possible to let the motion model vary over time, and adaptively update it based on data. Preliminary analysis shows that the adaptive multi-frame approach has the potential for yielding significant performance improvements over the conventional approach based on individual frames.

The Rayleigh Quotient Quadratic Correlation Filter (RQQCF) has been used to achieve very good performance for Automatic Target Detection/Recognition. The filter coefficients are obtained as the solution that maximizes a class separation metric, thus resulting in optimal performance. Recently, a transform domain approach was presented for ATR using the RQQCF called the Transform Domain RQQCF (TDRQQCF). The TDRQQCF considerably reduced the computational complexity and storage requirements, by compressing the target and clutter data used in designing the QCF. In addition, the TDRQQCF approach was able to produce larger responses when the filter was correlated with target and clutter images. This was achieved while maintaining the excellent recognition accuracy of the original spatial domain RQQCF algorithm. The computation of the RQQCF and the TDRQQCF involve the inverse of the term A₁ = R_x + R_y where R_x and R_y are the sample autocorrelation matrices for targets and clutter respectively. It can be conjectured that the TDRQQCF approach is equivalent to regularizing A₁. A common regularization approach involves performing the Eigenvalue Decomposition (EVD) of A₁, setting some small eigenvalues to zero, and then reconstructing A₁, which is now expected to be better conditioned. In this paper, this regularization approach is investigated, and compared to the TDRQQCF.

Mixed state or hybrid state space systems are useful tools for various problems in computer vision. These systems model complicated system dynamics as a mixture of inherently simple sub-systems, with an additional mechanism to switch between the sub-systems. This approach of modeling using simpler systems allows for ease in learning the parameters of the system and in solving the inference problem. In this paper, we study the use of such mixed state space systems for problems in recognition and behavior analysis in video sequences. We begin with a dynamical system formulation for recognition of faces from a video. This system is used to introduce the simultaneous tracking and recognition paradigm that allows for improved performance in both tracking and recognition. We extend this framework to design a second system for verification of vehicles across non-overlapping views using structural and textural fingerprints for characterizing the identity of the target. Finally, we show the use of such modeling for tracking and behavior analysis of bees from video.

This paper presents a novel analytical method measuring the impact of reference model accuracy on tactical, model-based automatic target acquisition (ATA) algorithm performance. Military tacticians are currently governed by various standards regarding quality requirements for georeferenced imagery and geospatial data sources. In some new generation systems, this imagery provides the basis for generating 3D reference models for input into model-based ATA systems. This paper analyses the criticality of this absolute coordinate accuracy requirement by assessing ATA algorithm performance using 3D reference models created from a variety of commercially available data sources, including aerial and terrestrial photography, and airborne laser scanner data. ATA algorithm performance is analysed using a software tool that uses a variety of open source techniques and image processing functions typically found in tactical, air-to-ground, model-based ATA systems. Each of the 3D reference models, extracted in a number of different areas of interest, was matched against a corresponding sequence of infrared video data. This provided a series of results, which were analysed as a function of both reference model accuracy and object selection and representation. Initial results indicate that the absolute accuracy of the reference models created for this research has a minor impact on ATA algorithm performance when compared with the impact of the content of the reference models, taking into account the complexity of the area of interest. This suggests that a wider array of data sources may be suitable for 3D reference model construction, than is currently accepted.

The paper considers the following problem: given a 3D model of a reference target and a sequence of images of a 3D scene, identify the object in the scene most likely to be the reference target and determine its current pose. Finding the best match in each frame independently of previous decisions is not optimal, since past information is ignored. Our solution concept uses a novel Bayesian framework for multi target tracking and object recognition to define and sequentially update the probability that the reference target is any one of the tracked objects. The approach is applied to problems of automatic lock-on and missile guidance using a laser radar seeker. Field trials have resulted in high target hit probabilities despite low resolution imagery and temporarily highly occluded targets.

Targeting and precision-guided munitions rely on precise image and video registration. Current approaches for geo-registration typically utilize registration algorithms that operate in two-dimensional (2D) transform spaces, in the absence of an underlying three-dimensional (3D) surface model. However, because of their two-dimensional motion assumptions, these algorithms place limitations on the types of imagery and collection geometries that can be used. Incorporating a 3D reference surface enables the use of 2D-to-3D registration algorithms and removes many such limitations. The author has previously demonstrated a fast 2D-to-3D registration algorithm for registration of live video to surface data extracted from medical images. The algorithm uses an illumination-tolerant gradient-descent based optimization to register a 2D image to 3D surface data in order to globally locate the camera's origin with respect to the 3D model. The rapid convergence of the algorithm is achieved through a reformulation of the optimization problem that allows many data elements to be re-used through multiple iterations. This paper details the extension of this algorithm to the more difficult problem of registering aerial imagery to terrain data.

Automatic detection of surface objects, like vessels, in a maritime environment from images is an important issue in naval surveillance. Two different approaches - gradient filter and background estimation - are presented in this paper and the test results on real data, both infrared as well as visible light images, are discussed. In the gradient approach, a gradient filter scans the sea-part of the image horizontally and vertically resulting in peaks at locations where the gradient exceeds a predefined local threshold. In the second approach, the background estimation, a polynomial model of the background is fitted locally to the seapart of the image. Using these polynomial background-estimators in the actual sea-analysis, objects are detected. In this paper the advantages and disadvantages of both approaches are discussed.

Edge features are often used in computer vision for image exploitation algorithms. A method to extract edge features that is robust to contrast change, translation, rotation, noise and scale change is presented. This method consists of the following steps: decompose the image into it's level set shapes, smooth the shapes, locate sections of the shape borders that have nearly constant curvature, and locate a key point based on these curve sections. The level sets are found using the Fast Level Set Transform (FLST). An affine invariant smoothing technique was then applied to the level set shape borders to reduce pixel effects and noise, and an intrinsic scale was estimated from the level set borders. The final step was key point location and scale estimation using the Helmholtz principle. These key points were found to be more resilient to large scale changes than the SIFT key points.

In this paper we introduce and test a new similarity measure for use in a template matching process for target detection and recognition. The measure has recently been developed for multi-modal registration of medical images and is known as phase mutual information (PMI). The key advantage of PMI is that it is invariant to lighting conditions, the ratio between foreground and background intensity and the level of background clutter, which is critical for target detection and recognition from the surveillance images acquired from various sensors. Several experiments were conducted using real and synthetic datasets to evaluate the performance of PMI when compared with a number of commonly used similarity measures including mean squared difference, gradient error and intensity mutual information. Our results show that PMI consistently provided the most accurate detection and recognition performance.

Humans are better at detecting targets in literal imagery than any known algorithm. Recent advances in modeling visual processes have resulted from f-MRI brain imaging with humans and the use of more invasive techniques with monkeys. There are four startling new discoveries. 1) The visual cortex does not simply process an incoming image. It constructs a physics based model of the image. 2) Coarse category classification and range-to-target are estimated quickly - possibly through the dorsal pathway of the visual cortex, combining rapid coarse processing of image data with expectations and goals. This data is then fed back to lower levels to resize the target and enhance the recognition process feeding forward through the ventral pathway. 3) Giant photosensitive retinal ganglion cells provide data for maintaining circadian rhythm (time-of-day) and modeling the physics of the light source. 4) Five filter types implemented by the neurons of the primary visual cortex have been determined. A computer model for automatic target detection has been developed based upon these recent discoveries. It uses an artificial neural network architecture with multiple feed-forward and feedback paths. Our implementation's efficiency derives from the observation that any 2-D filter kernel can be approximated by a sum of 2-D box functions. And, a 2-D box function easily decomposes into two 1-D box functions. Further efficiency is obtained by decomposing the largest neural filter into a high pass filter and a more sparsely sampled low pass filter.

The analysis of complex infrastructure from aerial imagery, for instance a detailed analysis of an airfield, requires the interpreter, besides to be familiar with the sensor's imaging characteristics, to have a detailed understanding of the infrastructure domain. The required domain knowledge includes knowledge about the processes and functions involved in the operation of the infrastructure, the potential objects used to provide those functions and their spatial and functional interrelations. Since it is not possible yet to provide reliable automatic object recognition (AOR) for the analysis of such complex scenes, we developed systems to support a human interpreter with either interactive approaches, able to assist the interpreter with previously acquired expert knowledge about the domain in question, or AOR methods, capable of detecting, recognizing or analyzing certain classes of objects for certain sensors. We believe, to achieve an optimal result at the end of an interpretation process in terms of efficiency and effectivity, it is essential to integrate both interactive and automatic approaches to image interpretation. In this paper we present an approach inspired by the advancing semantic web technology to represent domain knowledge, the capabilities of available AOR modules and the image parameters in an explicit way. This enables us to seamlessly extend an interactive image interpretation environment with AOR modules in a way that we can automatically select suitable AOR methods for the current subtask, focus them on an appropriate area of interest and reintegrate their results into the environment.

Three-dimensional (3D) Laser Detection and Ranging (LADAR) range data is being investigated for automatic target recognition applications. The spin-image provides a useful data representation for 3D point cloud data. In the spirit of recent work that shows ℓ₁-sparseness to be a useful data compression metric, we propose to use Nonnegative Matrix Factorization (NMF) to help find features that capture the salient information resident in the spin-image representation. NMF is a technique for decomposing nonnegative multivariate data into its 'parts', resulting in a compressed and usually sparse representation. As a surrogate for measured 3D LADAR data, we generate 3D point clouds from computer-aided-design models of two land targets, and we generate spin-images at multiple support scales. We select the support scale that provides the highest separability between the spin-image stacks from the two land targets. We then apply NMF to the spin-images at this support scale, and seek elements corresponding to meaningful parts of the land vehicles (e.g., a tank turret or truck wheels), that in a joint sense should provide significant discriminative capability. We measure the separability in the sparse NMF subspace. For measuring separability, we use the Henze-Penrose measure of multivariate distributional divergence.

In the intelligence community, aerial video has become one of the fastest growing data sources and it has been extensively used in intelligence, surveillance, reconnaissance, tactical and security applications. This paper presents a tracking approach to detect moving vehicles and person in such videos taken from aerial platform. In our approach, we combine the layer segmentation approach with background stabilization and post-tracking refinement to reliably detect small moving objects at the relatively low processing speed. For each individual moving object, a corresponding layer is created to maintain an independent appearance and motion model during the tracking process. After the online tracking process, we apply a post-tracking refinement process to link the track fragments into a long consistent track ID to further reduce false alarm and increase detection rate. Furthermore, a vehicle and person classifier is also integrated into the approach to identify the moving object categories. The classifier is based on image histogram of gradient (HOG), which is more reliable to illumination variation or camera automatic gain change. Finally, we report the results of our algorithms on a large scale of EO and IR data set collected from VIVID program, and the results show that our approach achieved a good and stable tracking performance on the data set that is more than eight hours.

Automatic target detection (ATD) is a very challenging problem for the Army in ground-to-ground scenarios using infrared (IR) sensors. I propose an ATD algorithm based on vector quantization (VQ). VQ is typically used for image compression where a codebook is created using the Linde Buzo Gray (LBG) algorithm from an example image. The codebook will be trained on clutter images containing no targets thus creating a clutter codebook. The idea is to encode and decode new images using the clutter codebook and calculate the VQ error. The error due solely to the compression will be approximately consistent across the image. In the areas that contain new objects in the scene (objects the codebook has not been trained on) we should see the consistent compression error plus an increased "non-training error" due to the fact that pixel blocks representing the new object are not included in the codebook. After the decoding process, areas in the image with large overall error will correlate to pixel blocks not in the codebook. The Kolomogorov-Smirnov distance is used to classify new objects from a reference clutter error distribution. The VQ algorithm trains on clutter so it will never have a problem with new targets like many "trained algorithms". The algorithm is run over a data set of images and the results show that the VQ detection algorithm performs as well as the Army benchmark algorithm.

In a FLIR image sequence, a target may disappear permanently or may reappear after some frames and crucial information such as direction, position and size related to the target are lost. If the target reappears at a later frame, it may not be tracked again because the 3D orientation, size and location of the target might be changed. We have proposed two methods: FKT-DCCF and FKT-PDCCF. We have implemented several tests by using DCCF and PDCCF when a target reappears to field-of-view. At the end of those test results we compared the DCCF and PDCCF methods by looking the percentage of correctly identified targets. Test results using both mid-wave and long-wave FLIR sequences (M1415, L1805, L1702, L1920, L19NS, and L1915) are incorporated to verify the effectiveness of the proposed algorithm.

Proposed distributed wavelet-based algorithms are a means to compress sensor data received at the nodes forming a wireless sensor network (WSN) by exchanging information between neighboring sensor nodes. Local collaboration among nodes compacts the measurements, yielding a reduced fused set with equivalent information at far fewer nodes. Nodes may be equipped with multiple sensor types, each capable of sensing distinct phenomena: thermal, humidity, chemical, voltage, or image signals with low or no frequency content as well as audio, seismic or video signals within defined frequency ranges. Compression of the multi-source data through wavelet-based methods, distributed at active nodes, reduces downstream processing and storage requirements along the paths to sink nodes; it also enables noise suppression and more energy-efficient query routing within the WSN. Targets are first detected by the multiple sensors; then wavelet compression and data fusion are applied to the target returns, followed by feature extraction from the reduced data; feature data are input to target recognition/classification routines; targets are tracked during their sojourns through the area monitored by the WSN. Algorithms to perform these tasks are implemented in a distributed manner, based on a partition of the WSN into clusters of nodes. In this work, a scheme of collaborative processing is applied for hierarchical data aggregation and decorrelation, based on the sensor data itself and any redundant information, enabled by a distributed, in-cluster wavelet transform with lifting that allows multiple levels of resolution. The wavelet-based compression algorithm significantly decreases RF bandwidth and other resource use in target processing tasks. Following wavelet compression, features are extracted. The objective of feature extraction is to maximize the probabilities of correct target classification based on multi-source sensor measurements, while minimizing the resource expenditures at participating nodes. Therefore, the feature-extraction method based on the Haar DWT is presented that employs a maximum-entropy measure to determine significant wavelet coefficients. Features are formed by calculating the energy of coefficients grouped around the competing clusters. A DWT-based feature extraction algorithm used for vehicle classification in WSNs can be enhanced by an added rule for selecting the optimal number of resolution levels to improve the correct classification rate and reduce energy consumption expended in local algorithm computations. Published field trial data for vehicular ground targets, measured with multiple sensor types, are used to evaluate the wavelet-assisted algorithms. Extracted features are used in established target recognition routines, e.g., the Bayesian minimum-error-rate classifier, to compare the effects on the classification performance of the wavelet compression. Simulations of feature sets and recognition routines at different resolution levels in target scenarios indicate the impact on classification rates, while formulas are provided to estimate reduction in resource use due to distributed compression.

We propose several fusion techniques in the design of a hybrid composite classification system. Our composite classifier taps into the strengths of two separate classification paradigms and examines various fusion methods for combining the two. The first classifier uses non-numeric features, similar to those found in syntactic pattern recognition, by exploiting the overall structure of the patterns themselves. The second method uses a more classical feature vector method that bins the patterns and uses the maximum values within each bin in developing the feature vector for each pattern. By using these two separate approaches, we explore conditions that allow the two techniques to be complementary in nature, thus improving, when fused, the overall performance of the classification system. We examine four seperate fusion techniques, the Basic Ensemble Method, the Probabilistic Neural Network, the Borda Count and the Bayesian Belief Network using a ten class problem in our experiments.

Small Unmanned Aerial Vehicles (UAVs) are increasingly being used in-theater to provide low-cost, low-profile aerial reconnaissance and surveillance capabilities. However, inherent platform limitations on size, weight, and power restrict the ability to provide sensors and communications which can present high-resolution imagery to the end-user. This paper discusses methods to alleviate this restriction by performing on-board pre-processing of high resolution images and downlinking the post-processed imagery. This has the added benefit of reducing the workload for a warfighter who is already heavily taxed by other duties.

During a previous technology programme, a simple landscape and complex target geometries were modelled and demonstrated in a COTS infrared (IR) simulation tool. A preliminary assessment of training-based ATR on real and synthetic imagery was performed, which was presented at SPIE D&S in 2005. The current technology programme has assessed model-based ATR on real and synthetic IR imagery for a 5-class case. Real IR imagery was recorded during a flight campaign. A complex landscape and complex targets were modelled and simulated in a wide variety of conditions in the IR simulation tool. A survey was conducted regarding the current state-of-the-art of model-based ATR approaches. Another survey concerning contour extraction methods for ATR was performed. The best ATR algorithms and contour extraction methods were selected from the survey results. These algorithms were implemented for a multi-class ATR case and adapted to work on the characteristics of IR imagery. The algorithms were benchmarked and compared on the simulated and recorded IR imagery using classical measures. A process for performance assessment of multi-class ATR methods was defined according to an ATR benchmarking concept developed by the German Fraunhofer Research Institute. The assessment was then conducted on the algorithms using a multi-class evaluation approach.