This PDF file contains the front matter associated with SPIE Proceedings Volume 8391, including the Title Page, Copyright information, Table of Contents, and the Conference Committee listing.

The detection, tracking, and classification of humans in video imagery is of obvious military and civilian importance. The problem is difficult under the best of circumstances. In infrared (IR) imagery, or any grayscale imagery, the problem is compounded by the lack of color cues. Sometimes, human detection in IR imagery can take advantage of the thermal difference between humans and background-but this difference is not robust. Varying environmental conditions regularly degrade the thermal contrast between humans and background. In difficult data, humans can be effectively camouflaged by their environment and standard feature detectors are unreliable. The research described here uses a hybrid approach toward human detection, tracking, and classification. The first is a feature-based correlated body parts detector. The second is a pseudo-Hough transform applied to the edge images of the video sequence. The third relies on an optical flow-based vector field transformation of the video sequence. This vector field permits a multidimensional application of the feature detectors initiated in the previous two methods. Then a multi-dimensional oriented Haar transform is applied to the vector field to further characterize potential detections. This transform also shows potential for distinguishing human behavior.

Forward Looking Infrared (FLIR) automatic target recognition (ATR) systems depend upon the capacity of the atmosphere to propagate thermal radiation over long distances. To date, not much research has been conducted on analyzing and mitigating the effects of the atmosphere on FLIR ATR performance, even though the atmosphere is often the limiting factor in long-range target detection and recognition. The atmosphere can also cause frame-to-frame inconsistencies in the scene, affecting the ability to detect and track moving targets. When image quality is limited by turbulence, increasing the aperture size or improving the focal plane array cannot improve ATR performance. Traditional single frame image enhancement does not solve the problem. A new approach is described for reducing the effects of turbulence. It is implemented under a lucky-region-imaging framework using short integration time and spatial domain processing. It is designed to preserve important target and scene structure. Unlike previous Fourier-based approaches originating from the astronomical community, this new approach is intended for real-time processing from a moving platform, with ground as the background. The system produces a video stream with minimal delay. A new video quality measure (VQM_turb) is presented for quantifying the success of turbulence mitigation on real data where no reference imagery is available. The VQM_turb is the core of the innovation because it allows a wide range of algorithms to be quantitatively compared. An algorithm can be chosen, and then tuned, to best-fit available processing power, latency requirements, scenarios and sensor characteristics.

Time series modeling is proposed for identification of targets whose images are not clearly seen. The model building takes into account air turbulence, precipitation, fog, smoke and other factors obscuring and distorting the image. The complex of library data (of images, etc.) serving as a basis for identification provides the deterministic part of the identification process, while the partial image features, distorted parts, irrelevant pieces and absence of particular features comprise the stochastic part of the target identification. The missing data approach is elaborated that helps the prediction process for the image creation or reconstruction. The results are provided.

In this paper, a robust automatic target recognition algorithm in FLIR imagery is proposed. Target is first segmented out from the background using wavelet transform. Segmentation process is accomplished by parametric Gabor wavelet transformation. Invariant features that belong to the target, which is segmented out from the background, are then extracted via moments. Higher-order moments, while providing better quality for identifying the image, are more sensitive to noise. A trade-off study is then performed on a few moments that provide effective performance. Bayes method is used for classification, using Mahalanobis distance as the Bayes' classifier. Results are assessed based on false alarm rates. The proposed method is shown to be robust against rotations, translations and scale effects. Moreover, it is shown to effectively perform under low-contrast objects in FLIR images. Performance comparisons are also performed on both GPU and CPU. Results indicate that GPU has superior performance over CPU.

Images obtained by airborne sensors are corrupted by distortions due to turbid and/or turbulent mediums between sensors and objects of interest on the ground. In this paper we present automated methods for compensating for this distortion by exploiting information available through use of multi-aperture passive imaging, and formation of composite bestpreserved enhanced imagery. Preliminary results of applying these methods on real and simulated data indicate their robustness in enhancing severely degraded imagery.

Vibration signatures sensed from distant vehicles using laser vibrometry systems provide valuable information that may be used to help identify key vehicle features such as engine type, engine speed, and number of cylinders. While developing algorithms to blindly extract the aforementioned features from a vehicle's vibration signature, it was shown that detection of engine speed and number of cylinders was more successful when utilizing a priori knowledge of the engine type (gas or diesel piston) and optimizing algorithms for each engine type. In practice, implementing different algorithms based on engine type first requires an algorithm to determine whether a vibration signature was produced by a gas piston or diesel piston engine. This paper provides a general overview of the observed differences between datasets from gas and diesel piston engines, and proceeds to detail the current method of differentiating between the two. To date, research has shown that basic signal processing techniques can be used to distinguish between gas and diesel vibration datasets with reasonable accuracy for piston engines of different configurations running at various speeds.

We present two methods to accurately estimate the location of an emitter from the received signal. The first of these methods is a variation of the standard cross-ambiguity function (CAF) for which we introduce a cross- spectral frequency estimation technique to replace the conventional methods based on the power spectrum. We demonstrate the use of the CSCAF in estimating the source location of an RF emission. The CSCAF is computed as the product of the complex-valued CAF and the conjugate of the time-delayed CAF. The magnitude of the CSCAF is the conventional CAF energy surface, and the argument of the CSCAF is the unquantized frequency difference of arrival (FDOA) computed as the phase of the CAF differentiated with respect to time. The advantage of the CSCAF is that it provides an extremely accurate estimate of FDOA. We demonstrate the use of the CSCAF in providing emitter location estimates that are superior to those provided by the conventional CAF. The second method presented provides a precision geolocation of the emitter from the signal received by a single moving receiver. This method matches the Doppler characteristics of the received signal to the Doppler characteristics estimated from the geometry of the transmitter and receiver. Both the CSCAF and the single receiver methods are enabled by cross-spectral frequency estimation methods that provide extremely accurate frequency estimation and tracking.

This paper utilizes game-theoretic principles in the automatic recognition of unknown radar targets. This study uses a non-cooperative matching game where pure strategies are associated with specific items to be matched, and agreement between possible hypotheses represents the payoff gained when playing a certain strategy against an opponent who is playing another strategy. The target recognition approach attempts to match scattering centers of an unknown target with those of library targets as competing strategies. The algorithm is tested using real radar data representing scattering from commercial aircraft models. Radar data of library targets at various azimuth positions are matched against an unknown radar target signature at a specific aspect angle. Computer simulations provide an estimate of the error rates in scenarios of additive Gaussian noise corrupting target signatures.

The paper presents new techniques and processing results for automatic segmentation, shape classification, generic pose estimation, and model-based identification of naval vessels in laser radar imagery. The special characteristics of focal plane array laser radar systems such as multiple reflections and intensity-dependent range measurements are incorporated into the algorithms. The proposed 3D model matching technique is probabilistic, based on the range error distribution, correspondence errors, the detection probability of potentially visible model points and false alarm errors. The match algorithm is robust against incomplete and inaccurate models, each model having been generated semi-automatically from a single range image. A classification accuracy of about 96% was attained, using a maritime database with over 8000 flash laser radar images of 146 ships at various ranges and orientations together with a model library of 46 vessels. Applications include military maritime reconnaissance, coastal surveillance, harbor security and anti-piracy operations.

We have previously developed a feature extraction process for propagation-invariant classification of a target from its propagated sonar backscatter. The features are invariant to the frequency dependent propagation effects of absorption and dispersion, for range-independent channels. Simulations have shown that these features lose their effectiveness when applied to waves propagating in a range-dependent environment. In this paper we extend our previous approach to obtain invariant features for classification in range-dependent environments. Numerical simulations are presented for the classification of two shells from their acoustic backscatter propagating in an ideal wedge.

We present two methods to obtain a probability distribution from its moments. In the first method one chooses a set of orthogonal polynomials with corresponding weighting function. In the second method one chooses a starting distribution and uses that to construct orthogonal polynomials. We give a number of examples.

We have calculated the raw and central moments of a noise model where the noise is the sum of elementary signals which are time and space dependent. We show that the scintillation index is a constant plus a correction term that goes as 1/N where N is the number of elementary signals. We study the correction term for a specific elementary signal.

Automated pattern recognition has been around for several decades. Generally we have been successful at tackling a large variety of technical problems. This paper presents the challenges inherent in wide area motion imagery, which is problematic to the normal pattern recognition process and associated pattern recognition systems. This paper describes persistent wide area motion imagery, its role as a manifold for overlaying episodic sensors of various modalities to present a better view of activity to an analyst. An underlying framework, SPADE, is introduced and a layered sensing viewer, Pursuer, is also presented to demonstrate the utility of creating a unified view of the sensing world to an analyst.

The ability to recognize a target in an image is an important problem for machine vision, surveillance systems, and military weapons. There are many "solutions" to an automatic target recognition (ATR) problem proposed by practitioners. Often the definition of the problem leads to multiple solutions due to the incompleteness of the definition. Solutions are also made approximate due to resource limitations. Issues concerning "best" solution and solution performance are very open issues, since problem definitions and solutions are ill-defined. Indeed from information based physical measurement theory such as found in the Minimum Description Length (MDL) the exact solution is intractable1. Generating some clarity in defining problems on restricted sets seems an appropriate approach for improving this vagueness in ATR definitions and solutions. Given that a one to one relationship between a physical system and the MDL exists, then this uniqueness allows that a solution can be defined by its description and a norm assigned to that description. Moreover, the solution can be characterized by a set of metrics that are based on the algorithmic information of the physical measurements. The MDL, however, is not a constructive theory, but solutions can be defined by concise problem descriptions. This limits the scope of the problem and we will take this approach here. The paper will start with a definition of an ATR problem followed by our proposal of a descriptive solution using a union of subspaces model of images as described below based on Lu and Do². This solution uses the concept of informative representations3 implicitly which we review briefly. Then we will present some metrics to be used to characterize the solution(s) which we will demonstrate by a simple example. In the discussions following the example we will suggest how this fits in the context of present and future work.

Techniques such as SIFT and SURF facilitate efficient and robust image processing operations through the use of sparse and compact spatial feature descriptors and show much potential for defence and security applications. This paper considers the extension of such techniques to include information from the temporal domain, to improve utility in applications involving moving imagery within video data. In particular, the paper demonstrates how spatio-temporal descriptors can be used very effectively as the basis of a target tracking system and as target discriminators which can distinguish between bipeds and quadrupeds. Results using sequences of video imagery of walking humans and dogs are presented, and the relative merits of the approach are discussed.

We present a general purpose, inherently robust system for object representation and recognition. The system is model-based and knowledge-based, with knowledge derived from analysis of objects and images, unlike many of the current methods which rely on generic statistical inference. This knowledge is intrinsic to the objects themselves, based on geometric and semantic relations among objects. Therefor the system is insensitive to external interferences such as viewpoint changes (scale, pose etc.), illumination changes, occlusion, shadows, sensor noise etc. It also handles variability in the object itself, e.g. articulation or camouflage. We represent all available models in a graph containing two independent but interlocking hierarchies. One of these intrinsic hierarchies is based on parts, e.g. a truck has a cabin, a trunk, wheels etc. The other hierarchy we call the "Level of Abstraction (LOA), e.g. a vehicle is more abstract than a truck, a rectangle is more abstract than a door. This enables us to represent and recognize generic objects just as easily as specific ones. A new algorithm for traversing our graph, combining the advantages of both top-down and bottom-up strategies, has been implemented.

We present a technique for small watercraft detection in a littoral environment characterized by multiple targets and both land- and sea-based clutter. The detector correlates a tailored wavelet model trained from previous imagery with newly acquired scenes. An optimization routine is used to learn a wavelet signal model that improves the average probability of detection for a xed false alarm rate on an ensemble of training images. The resulting wavelet is shown to improve detection on a previously unseen set of test images. Performance is quantied with ROC curves.

Efficient moving object tracking requires near flawless detection results to establish correct correspondences between frames. This is especially true in the defense sector where accuracy and speed are critical factors of success. However, problems such as camera motion, lighting and weather changes, texture variation and inter-object occlusions result in misdetections or false positive detections which in turn, lead to broken tracks. In this paper, we propose to use background subtraction and an optimized version of Horn & Schunk's optical flow algorithm in order to boost detection response. We use the frame differencing method, followed by morphological operations to show that it works in many scenarios and the optimized optical flow technique serves to complement the detector results. The Horn & Schunk's method yields color-coded motion vectors for each frame pixel. To segment the moving regions in the frame, we apply color thresholding to distinguish the blobs. Next, we extract appearance-based features from the detected object and establish the correspondences between objects' features, in our case, the object's centroid. We have used the Euclidean distance measure to compute the minimum distances between the centroids. The centroids are matched by using Hungarian algorithm, thus obtaining point correspondences. The Hungarian algorithm's output matrix dictates the objects' associations with each other. We have tested the algorithm to detect people in corridor, mall and field sequences and our early results with an accuracy of 86.4% indicate that this system has the ability to detect and track objects in video sequences robustly.

The set of orthogonal eigen-vectors built via principal component analysis (PCA), while very effective for com- pression, can often lead to loss of crucial discriminative information in signals. In this work, we build a new basis set using synthetic aperture radar (SAR) target images via non-negative matrix approximations (NNMAs). Owing to the underlying physics, we expect a non-negative basis and an accompanying non-negative coecient set to be a more accurate generative model for SAR proles than the PCA basis which lacks direct physical interpretation. The NNMA basis vectors while not orthogonal capture discriminative local components of SAR target images. We test the merits of the NNMA basis representation for the problem of automatic target recognition using SAR images with a support vector machine (SVM) classier. Experiments on the benchmark MSTAR database reveal the merits of basis selection techniques that can model imaging physics more closely and can capture inter-class variability, in addition to identifying a trade-off between classication performance and availability of training.

Automatic object recognition capabilities are traditionally tuned to exploit the specific sensing modality they were designed to. Their successes (and shortcomings) are tied to object segmentation from the background, they typically require highly skilled personnel to train them, and they become cumbersome with the introduction of new objects. In this paper we describe a sensor independent algorithm based on the biologically inspired technology of map seeking circuits (MSC) which overcomes many of these obstacles. In particular, the MSC concept offers transparency in object recognition from a common interface to all sensor types, analogous to a USB device. It also provides a common core framework that is independent of the sensor and expandable to support high dimensionality decision spaces. Ease in training is assured by using commercially available 3D models from the video game community. The search time remains linear no matter how many objects are introduced, ensuring rapid object recognition. Here, we report results of an MSC algorithm applied to object recognition and pose estimation from high range resolution radar (1D), electrooptical imagery (2D), and LIDAR point clouds (3D) separately. By abstracting the sensor phenomenology from the underlying a prior knowledge base, MSC shows promise as an easily adaptable tool for incorporating additional sensor inputs.

Compressive imaging is an emerging field which allows one to acquire far fewer measurements of a scene than a standard pixel array and still retain the information contained in the scene. One can use these measurements to reconstruct the original image or even a processed version of the image. Recent work in compressive imaging from random convolutions is extended by relaxing some model assumptions and introducing the latest sparse reconstruction algorithms. We then compare image reconstruction quality of various convolution mask sizes, compression ratios, and reconstruction algorithms. We also expand the algorithm to derive a pattern recognition system which operates of a compressively sensed measurement stream. The developed compressive pattern recognition system reconstructions the detections map of the scene without the intermediate step of image reconstruction. A case study is presented where pattern recognition performance of this compressive system is compared against a full resolution image.

Covariance estimation is a key step in many target detection algorithms. To distinguish target from background requires that the background be well-characterized. This applies to targets ranging from the precisely known chemical signatures of gaseous plumes to the wholly unspecified signals that are sought by anomaly detectors. When the background is modelled by a (global or local) Gaussian or other elliptically contoured distribution (such as Laplacian or multivariate-t), a covariance matrix must be estimated. The standard sample covariance overfits the data, and when the training sample size is small, the target detection performance suffers. Shrinkage addresses the problem of overfitting that inevitably arises when a high-dimensional model is fit from a small dataset. In place of the (overfit) sample covariance matrix, a linear combination of that covariance with a fixed matrix is employed. The fixed matrix might be the identity, the diagonal elements of the sample covariance, or some other underfit estimator. The idea is that the combination of an overfit with an underfit estimator can lead to a well-fit estimator. The coefficient that does this combining, called the shrinkage parameter, is generally estimated by some kind of cross-validation approach, but direct cross-validation can be computationally expensive. This paper extends an approach suggested by Hoffbeck and Landgrebe, and presents efficient approximations of the leave-one-out cross-validation (LOOC) estimate of the shrinkage parameter used in estimating the covariance matrix from a limited sample of data.

Target detection and classification is very crucial in wireless sensor network for outdoor security applications. This paper presents a novel concept to detect and further discriminate dynamic objects (persons and vehicles) in WSN using geophones that work as dynamic presence detectors in a given field of interest. The basic concept to detect and classify a target lies in the idea to consider an individual footstep of a person and/or motion of a vehicle detects as an event in the detection range of the geophone. In an on-going project, the design of a wireless geophone sensor node is implemented. This design is based on high processing Gumstix Overo Fire Computer-On-Module that is used for high performance and has built-in wireless capabilities. A high resolution analog-to-digital converter is integrated to the module to acquire data from geophone. The raw data is processed on the sensor node to detect and classify a target. The adaptive wavelet denoising algorithm is applied in real-time to extract a target signal from the real noisy environments. This algorithm adjusts the threshold based on the energy of the wavelet series coefficients. Timestamps of the events are extracted using event detection method. These events are used to classify a target on a node level.

Target detection is an important research content in hyperspectral remote sensing technology, which is widely used in securities and defenses. Nowadays, many target detection algorithm have been proposed. One of the key evaluation indicators of these algorithms performance is false-alarm rate. The feature-level fusion of different target detection results is a simple and effective method to reduce false-alarm rate. But the different value ranges of different algorithms bring difficulties for data fusion. This paper proposed a feature-level fusion method based on RXD detector, which is to integrate multiple target detection results into a multi-bands image, and fuse detection results using principal theory of abnormal detection. Experiments revealed that, this method is not restricted by the quantity of target detection algorithms and not influenced by different value ranges of different algorithms, which can reduce false-alarm rate effectively.