Proceedings Volume 7335

Automatic Target Recognition XIX

cover
Proceedings Volume 7335

Automatic Target Recognition XIX

View the digital version of this volume at SPIE Digital Libarary.

Volume Details

Date Published: 28 April 2009
Contents: 11 Sessions, 37 Papers, 0 Presentations
Conference: SPIE Defense, Security, and Sensing 2009
Volume Number: 7335

Table of Contents

icon_mobile_dropdown

Table of Contents

All links to SPIE Proceedings will open in the SPIE Digital Library. external link icon
View Session icon_mobile_dropdown
  • Front Matter: Volume 7335
  • Advanced Concepts/Algorithms in ATR I
  • Advanced Concepts/Algorithms in ATR II
  • Multi- and Hyperspectral Processing in ATR
  • Radar Processing for ATR
  • Performance Evaluation Issues in ATR I
  • Performance Evaluation Issues in ATR II
  • Advanced Concepts/Algorithms in ATR III
  • Advanced Concepts/Algorithms in ATR IV
  • Advanced Concepts/Algorithms in ATR V
  • Polarimetric and Infrared Processing for ATR
Front Matter: Volume 7335
icon_mobile_dropdown
Front Matter: Volume 7335
This PDF file contains the front matter associated with SPIE Proceedings Volume 7335, including the Title Page, Copyright information, Tabe of Contents, Introduction (if any), and the Conference Committee listing.
Advanced Concepts/Algorithms in ATR I
icon_mobile_dropdown
High speed automatic target recognition based on anisotropic diffusion and discrete cosine transform
Automatic Target Recognition is one of the most challenging and important requirements in the 21st century battlefield. Developing an algorithm which is complex enough to recognize targets and simple enough to run in real time is a challenging problem. Recognizing different targets with different size, orientation and illumination variations increase the complexity of the problem. This paper proposes a recognition approach, which tries to recognize targets fast and correctly providing robust performance. The proposed algorithm is based on anisotropic diffusion and edge detection for image segmentation and discrete cosine transform (DCT) for image classification. First, difference between target and background is increased by using the anisotropic diffusion filter. In this method, diffusion continues over low contrast pixels, decreasing the difference between smooth regions. On the other hand, diffusion stops over high contrast pixels such that the sharper boundaries are preserved. Anisotropic diffusion method controls the directions of diffusion by an error function which separates low-contrast and high-contrast neighbor pixels. Instead of using partial differential equations or robust statistical equations as an error function, a simple threshold is used to decrease iteration number and operation time. Secondly, possible targets are segmented by using "Canny's edge detection" algorithm and "connected component labeling" algorithm. Finally, possible targets and target database dimensions are reduced and compared by DCT algorithm. In order to minimize the effect of illumination variations, low frequency coefficients aren't used in this comparative study. The proposed algorithm is then tested using example pictures, and is able to find targets in less than a second.
Cellular automata enabling novel fast shape recognition for muon tomography
Holger M. Jaenisch, James W. Handley, Kristina L. Jaenisch, et al.
We present a novel and detailed algorithm for enabling passive muon tomography systems to be used for 3-D threat object recognition in real-time. Our method makes use of characteristic changes of the Hamming distance curve derived from Cellular Automata rules converted into a novel Data Model form. We show that fragmented and noisy shape images can be adequately processed and recognized without resorting to morphological or traditional template matching approaches. The approach is general and has utility in other target/shape recognition and imaging applications.
Transformation of time dependent probability distributions
Juraj Tekel, Leon Cohen
Given a distribution of position at time zero, P(x, 0), one obtains the distribution at a subsequent time t, P(x, t), by solving the appropriate evolution equation. Often this is a very difficult problem. However, sometimes, it is relativity easy to obtain the exact time dependent low order moments. We present methods to approximate P(x, t) using the initial probability distribution and exact low order time dependent moments.
Advanced Concepts/Algorithms in ATR II
icon_mobile_dropdown
Contextual object understanding through geospatial analysis and reasoning (COUGAR)
Joel Douglas, Matthew Antone, James Coggins, et al.
Military operations in urban areas often require detailed knowledge of the location and identity of commonly occurring objects and spatial features. The ability to rapidly acquire and reason over urban scenes is critically important to such tasks as mission and route planning, visibility prediction, communications simulation, target recognition, and inference of higher-level form and function. Under DARPA's Urban Reasoning and Geospatial ExploitatioN Technology (URGENT) Program, the BAE Systems team has developed a system that combines a suite of complementary feature extraction and matching algorithms with higher-level inference and contextual reasoning to detect, segment, and classify urban entities of interest in a fully automated fashion. Our system operates solely on colored 3D point clouds, and considers object categories with a wide range of specificity (fire hydrants, windows, parking lots), scale (street lights, roads, buildings, forests), and shape (compact shapes, extended regions, terrain). As no single method can recognize the diverse set of categories under consideration, we have integrated multiple state-of-the-art technologies that couple hierarchical associative reasoning with robust computer vision and machine learning techniques. Our solution leverages contextual cues and evidence propagation from features to objects to scenes in order to exploit the combined descriptive power of 3D shape, appearance, and learned inter-object spatial relationships. The result is a set of tools designed to significantly enhance the productivity of analysts in exploiting emerging 3D data sources.
Moments of a wave propagating with dispersion and damping
Greg Okopal, Patrick Loughlin
When a wave propagates in a medium with dispersion and damping, different frequencies propagate at different velocities and are attenuated at different rates. Accordingly, the wave changes as it propagates. These propagation effects can negatively impact automatic classification, since what is observed changes from location to location. We examine various moments of a wave, such as duration and bandwidth, which are often used as features for classification, and quantify the effects of dispersion and damping on these moments. We also identify moment-like features that are invariant to dispersion and damping, and thus may offer advantages over ordinary moments as features for classification.
Comparison of kernel based PDF estimation methods
David E. Freund, Philippe Burlina, Amit Banerjee, et al.
There are a number of challenging estimation, tracking, and decision theoretic problems that require the estimation of Probability Density Functions (PDFs). When using a traditional parametric approach, the functional model of the PDF is assumed to be known. However, these models often do not capture the complexity of the underlying distribution. Furthermore, the problems of validating the model and estimating its parameters are often complicated by the sparsity of prior examples. The need for exemplars grows exponentially with the dimension of the feature space. These methods may yield PDFs that do not generalize well to unseen data because these tend to overfit or underfit the training exemplars. We investigate and compare alternate approaches for estimating a PDF and consider instead kernel based estimation methods which generalize the Parzen estimator and use a Linear Mixture of Kernels (LMK) model. The methods reported here are derived from machine learning methods such as the Support Vector Machines and the Relevance Vector Machines. These PDF estimators provide the following benefits: (a) they are data driven; (b) they do not overfit the data and consequently have good generalization properties; (c) they can accommodate highly irregular and multi-modal data distributions; (d) they provide a sparse and succinct description of the underlying data which leads to efficient computation and communication. Comparative experimental results are provided illustrating these properties using simulated Mixture of Gaussian-distributed data.
Automatic building identification under bomb damage conditions
Robert Woodley, Warren Noll, Joseph Barker, et al.
Given the vast amount of image intelligence utilized in support of planning and executing military operations, a passive automated image processing capability for target identification is urgently required. Furthermore, transmitting large image streams from remote locations would quickly use available band width (BW) precipitating the need for processing to occur at the sensor location. This paper addresses the problem of automatic target recognition for battle damage assessment (BDA). We utilize an Adaptive Resonance Theory approach to cluster templates of target buildings. The results show that the network successfully classifies targets from non-targets in a virtual test bed environment.
Multi- and Hyperspectral Processing in ATR
icon_mobile_dropdown
3D multi-view passive sensing and visualization using randomly distributed sensors
Three dimensional (3D) passive imaging systems are proven to be effective in a number of applications including Automatic Target Recognition (ATR). Such systems are traditionally designed around a regular, fixed grid of pickup locations such as lenslet arrays - a constraint that can not always be met for certain applications. With the recent advancements in this area, many applications call for more generic form of 3D imaging. Here, we overview out work in the area of multi-perspective imaging based on randomly distributed passive sensing. In particular, we propose a passive 3D imaging and visualization system with multiple view acquisitions from randomly distributed sensors. This method can further extend the applications of passive 3D imaging systems to areas such as 3D aerial imaging, collaborative imaging and etc. We discuss some of implications for improving performance of ATR algorithms.
Wavelet-based hyperspectral target detection using spectral fringe-adjusted joint transform correlation
Wesam A. Sakla, Adel A. Sakla, M. S. Alam
Recently, the 1-D spectral fringe-adjusted joint transform correlation (SFJTC) technique has been combined with the discrete wavelet transform (DWT) as an effective means for providing robust target detection in hyperspectral imagery. This paper expands upon earlier work that demonstrates the utility of the DWT in conjunction with SFJTC for detection. We show that using selected DWT coefficients at a given decomposition level can significantly improve the ROC curve behavior of the detection process in comparison to using the original hyperspectral signatures. The DWT coefficients that are selected for detection are based on a supervised training process that uses the pure target signature and randomly selected samples from the scene. We illustrate this by conducting experiments on two different hyperspectral scenes containing varying amounts of simulated noise. Results show that use of the selected DWT coefficients significantly improves the ROC curve detection behavior in the presence of noise.
A support vector data description approach to target detection in hyperspectral imagery
Wesam A. Sakla, Adel A. Sakla, Andrew Chan
Spectral variability remains a challenging problem for target detection and classification in hyperspectral imagery (HSI). In this paper, we have applied the nonlinear support vector data description (SVDD) to perform full-pixel target detection. Using a pure target signature, we have developed a novel pattern recognition (PR) algorithm to train an SVDD to characterize the target class. We have inserted target signatures into an urban hyperspectral (HS) scene with varying levels of spectral variability to explore the performance of the proposed SVDD target detector in different scenarios. The proposed approach makes no assumptions regarding the underlying distribution of the scene data as do traditional statistical detectors such as the matched filter (MF). Detection results in the form of confusion matrices and receiver-operating- characteristic (ROC) curves demonstrate that the proposed SVDD-based algorithm is highly accurate and yields higher true positive rates (TPR) and lower false positive rates (FPR) than the MF.
A general purpose adaptive approach to image classification, automatic target detection, and recognition for multispectral imagery
Automatic target detection and recognition (ATD/ATR) can be considered one of the most sought after goals in image exploitation. There are no shortage of "good" algorithms in ATD/ATR on paper, yet the problem remains that an algorithm cannot be applied directly to different scenario and expect a similar success rate. One can attribute the difficulties in ATD/ATR to the ambiguity of the definition of "target", and the specific choice of image data and parameters in the design of the algorithms. We propose a general purpose approach to the problem in that we do not specify what a target is, except that it will be chosen by a user from a number of detected anomalies at the end of the classification cycle. At this time, a user can specify a number of attributes to be associated with a candidate target. There is a learning phase where the algorithm and the discriminating parameters are tuned based on the characteristics of the image data and the classification methods. There are a number of attributes associated with a target, both in spectral and spatial values, which can be set by a user. The number of bands used for input can be varied; however it is limited to three to seven bands at this point. Target recognition is achieved when a target candidate has a passing figure of merit, which again is defined by the user. It is hoped that this approach can provide a framework of ATD/ATR with greatest flexibility in algorithm re-use.
An interactive graphical performance analysis tool for hyperspectral exploitation evaluations
The nature of hyperspectral exploitation systems is such that a set of spectral imagery - and possibly a priori information such as a supplied library of target spectral signatures - is ingested into an algorithm and a series of responses is output. These responses must be scored for their accuracy against known target locations in the image set, from which algorithm performance is then determined. We propose, implement, and demonstrate a new environment for visualizing this process, which will aid not only the evaluator but also the algorithm developer in better understanding, characterizing, and improving system performance, be it that of an anomaly detection, change detection, or material identification algorithm.
Hyperspectral target detection in noisy environment using wavelet filter and correlation based detector
In this paper, we propose an algorithm for detecting man made targets in hyperspectral imagery using correlation based detection after wavelet domain filtering. In the proposed method, each spectral pixel in noisy hyperspectral data cube is filtered by wavelet domain filtering. Wavelet domain filtering looks at every spectral pixel as noisy signal and filter out noise through wavelet shrinkage based method. Then correlation between the provided target spectral signature and spectral signal from data cube is calculated. The algorithm scans each pixel in data cube then calculates correlation with target signature. The process yields correlation image. Applying threshold operation for correlation image provides detection image. The detection performance of the algorithm is tested with several hyperspectral datasets. Using ROC analysis and comparing with ground truth image, it is observed that wavelet based filtering provides better detection performance for noisy data. The simulation results indicate that the proposed algorithm efficiently detects object of interest in all datasets.
Multi-look fusion identification: a paradigm shift from quality to quantity in data samples
A multi-look identification method known as score-level fusion is found to be capable of achieving very high identification accuracy, even when low quality target signatures are used. Analysis using measured ground vehicle radar signatures has shown that a 97% correct identification rate can be achieved using this multi-look fusion method; in contrast, only a 37% accuracy rate is obtained when single target signature input is used. The results suggest that quantity can be used to replace quality of the target data in improving identification accuracy. With the advent of sensor technology, a large amount of target signatures of marginal quality can be captured routinely. This quantity over quality approach allows maximum exploitation of the available data to improve the target identification performance and this could have the potential of being developed into a disruptive technology.
Radar Processing for ATR
icon_mobile_dropdown
Probing waveform synthesis and receive filter design for active sensing systems
William Roberts, Hao He, Xing Tan, et al.
Probing waveform synthesis and receive filter design play crucial roles in achievable performance for active sensing applications, including radar, sonar, and medical imaging. We focus herein on conventional single-input single-output (SISO) radar systems. A flexible receive filter design approach, at the costs of lower signal-to-noise ratio (SNR) and higher computational complexity, can be used to compensate for missing features of the probing waveforms. A well synthesized waveform, meaning one with good autocorrelation properties, can reduce computational burden at the receiver and improve performance. Herein, we will highlight the interplay between waveform synthesis and receiver design. We will review a novel, cyclic approach to waveform design, and then compare the merit factors of these waveforms to other well-known sequences. In our comparisons, we will consider chirp, Frank, Golomb, and P4 sequences. Furthermore, we will overview several advanced techniques for receiver design, including data-independent instrumental variables (IV) filters, a data-adaptive iterative adaptive approach (IAA), and a data-adaptive Sparse Bayesian Learning (SBL) algorithm. We will show how these designs can significantly outperform conventional matched filter (MF) techniques for range compression as well as for range-Doppler imaging.
On the relationship between the generalized likelihood ratio test and backprojection for synthetic aperture radar imaging
In synthetic aperture radar (SAR) imaging, a scene of interest is illuminated by electromagnetic waves. The aim is to reconstruct an image of the scene from the measurement of the scattered waves using airborne antenna(s). There are many imaging systems which are built upon this notion such as mono-static SAR, bi-static SAR, and hitchhiker SAR. For these modalities, there are analytic reconstruction algorithms based on backprojection. Backprojection-based algorithms have the advantage of putting the visible edges of the scene at the right location and orientation in the reconstructed images. On the other hand, there is also a SAR imaging method based on the generalized likelihood-ratio test (GLRT). In particular we consider the problem of detecting a target at an unknown location. In the GLRT, the presence of a target in the scene is determined based on the likelihood-ratio test. Since the location of the target is not known, the GLRT test statistic is calculated for each position in the scene and the location corresponding to the maximum test statistic indicates the location of a potential target. In this paper, we show that the backprojection-based analytic reconstruction methods include as a special case the GLRT method. We show that the GLRT test statistic is related to the reflectivity of the scene when a backprojection-based reconstruction algorithm is used.
Time-frequency representations and operators
We discuss the issues that arise in developing quasi-distributions for arbitrary operators in contrast to the usual case of time and frequency. The arbitrary operator case has many mathematical and physical challenges that have not been solved. We also discuss the connection with differential equations and pseudo-differential operators. In regard to differential equations we argue that the proper generalization of the constant coefficient case is not the variable coefficients case but an equation where the coefficients are kept constant and the differential operator is replaced by a Hermitian operator.
On the airworthiness approval of a SAR ATR system
Dieter Willersinn, Uwe Jäger, Herbert Schlatt, et al.
A manned platform is to be equipped with a Synthetic Aperture Radar (SAR) based Automatic Target Recognition (ATR) system for precision targeting. The platform's airworthiness has to be approved including the ATR system, i.e. the ATR system needs to be qualified appropriately. Part of the airworthiness approval is a hazard analysis. In general, this is carried out to make sure that the probability of a fatal error in one hour of flight is 10-9 or lower. To date, error probabilities of a SAR-based ATR system, i.e. error probabilities of detection and classification, must be assumed to lie above 10-9 per hour. This is one reason why existing rules of engagement demand "Man-in-the loop", i.e. to display the result of the ATR system to the pilot. Components to the ATR system are consequently a Synthetic Aperture Radar (SAR) sensor an Automatic Target Recognition (ATR) SAR image processing unit, and a Human Machine Interface (HMI) to the pilot. The aim of the work reported in this contribution was to identify those performance features of the thus defined ATR system that are relevant to airworthiness approval, and to define the procedures to determine the feature values. The paper contains the analysis of a reference case of an airworthiness-approved technical system with an error probability above 10-9 per hour and a result display to the pilot. In the light of the analysis results, it concludes with an outlook to the airworthiness approval of the ATR system.
Performance Evaluation Issues in ATR I
icon_mobile_dropdown
Image quality and performance modeling for automated target detection
John M. Irvine, Eric Nelson
Several methods have been developed for quantifying the information potential of imagery exploited by a human observer. The National Imagery Interpretability Ratings Scale (NIIRS) has proven to be a useful standard for intelligence, surveillance, and reconnaissance (ISR) applications. A comparable standard for automated information extraction would be useful for a variety of applications, including tasking and collection management. This paper examines the applicability of NIIRS to automated exploitation methods. In particular, we compare image-based estimates of the NIIRS to observed performance of an automated target detection (ATD) algorithm. In addition, we examine other image metrics and their relationship to ATD performance. The findings indicate that NIIRS is not a good predictor of ATD performance, but methods that quantify the complexity of the clutter hold promise.
Performance study on point target detection using super-resolution reconstruction
When bright moving objects are viewed with an electro-optical system at very long range, they will appear as small slightly blurred moving points in the recorded image sequence. Detection of point targets is seriously hampered by structure in the background, temporal noise and aliasing artifacts due to undersampling by the infrared (IR) sensor. Usually, the first step of point target detection is to suppress the clutter of the stationary background in the image. This clutter suppression step should remove the information of the static background while preserving the target signal energy. Recently we proposed to use super-resolution reconstruction (SR) in the background suppression step. This has three advantages: a better prediction of the aliasing contribution allows a better clutter reduction, the resulting temporal noise is lower and the point target energy is better preserved. In this paper the performance of the point target detection based on super-resolution reconstruction (SR) is evaluated. We compare the use of robust versus non robust SR reconstruction and evaluate the effect of regularization. Both of these effects are influenced by the number of frames used for the SR reconstruction and the apparent motion of the point target. We found that SR improves the detection efficiency, that robust SR outperforms non-robust SR, and that regularization decreases the detection performance. Therefore, for point target detection one can best use a robust SR algorithm with little or no regularization.
Performance Evaluation Issues in ATR II
icon_mobile_dropdown
Error estimation procedure for large dimensionality data with small sample sizes
Using multivariate data analysis to estimate the classification error rates and separability between sets of data samples is a useful tool for understanding the characteristics of data sets. By understanding the classifiability and separability of the data, one can better direct the appropriate resources and effort to achieve the desired performance. The following report describes our procedure for estimating the separability of given data sets. The multivariate tools described in this paper include calculating the intrinsic dimensionality estimates, Bayes error estimates, and the Friedman-Rafsky tests. These analysis techniques are based on previous work used to evaluate data for synthetic aperture radar (SAR) automatic target recognition (ATR), but the current work is unique in the methods used to analyze large dimensionality sets with a small number of samples. The results of this report show that our procedure can quantitatively measure the performance between two data sets in both the measure and feature space with the Bayes error estimator procedure and the Friedman- Rafsky test, respectively. Our procedure, which included the error estimation and Friedman-Rafsky test, is used to evaluate SAR data but can be used as effective ways to measure the classifiability of many other multidimensional data sets.
Determining training data requirements for template based normalized cross correlation
Peter Knee, Lee Montagnino, Shawn Halversen, et al.
In this paper, we investigate the effect of increasingly sparse training data sets on target classification performance using a template-based classifier. An often used method of template creation employs averaging of multiple target training chips for a predefined coverage swath. The inclusion of too many training chips results in a blurring of the predominant scatterers while averaging of too few training chips results in poor edge resolution. We use the public MSTAR data set to show that using all appropriate images for each template may not result in the best ATR performance. We successfully demonstrate the ability to reduce training data collection requirements by requiring fewer training chips per template.
Automated rapid training of ATR algorithms
Jonah McBride, Jessica Lowell, Magnús Snorrason, et al.
Computer vision methods, such as automatic target recognition (ATR) techniques, have the potential to improve the accuracy of military systems for weapon deployment and targeting, resulting in greater utility and reduced collateral damage. A major challenge, however, is training the ATR algorithm to the specific environment and mission. Because of the wide range of operating conditions encountered in practice, advanced training based on a pre-selected training set may not provide the robust performance needed. Training on a mission-specific image set is a promising approach, but requires rapid selection of a small, but highly representative training set to support time-critical operations. To remedy these problems and make short-notice seeker missions a reality, we developed Learning and Mining using Bagged Augmented Decision Trees (LAMBAST). LAMBAST examines large databases and extracts sparse, representative subsets of target and clutter samples of interest. For data mining, LAMBAST uses a variant of decision trees, called random decision trees (RDTs). This approach guards against overfitting and can incorporate novel, mission-specific data after initial training via perpetual learning. We augment these trees with a distribution modeling component that eliminates redundant information, ignores misrepresentative class distributions in the database, and stops training when decision boundaries are sufficiently sampled. These augmented random decision trees enable fast investigation of multiple images to train a reliable, mission-specific ATR. This paper presents the augmented random decision tree framework, develops the sampling procedure for efficient construction of the sample, and illustrates the procedure using relevant examples.
Advanced Concepts/Algorithms in ATR III
icon_mobile_dropdown
Uncertain geometry: a new approach to modeling for recognition
Over the last several years, a new representation for geometry has been developed, based on a 3-d probability distribution of surface position and appearance. This representation can be constructed from multiple images, using both still and video data. The probability for 3-d surface position is estimated in an on-line algorithm using Bayesian inference. The probability of a point belonging to a surface is updated as to its success in accounting for the intensity of the current image at the projected image location of the point. A Gaussian mixture is used to model image appearance. This update process can be proved to converge under relatively general conditions that are consistent with aerial imagery. There are no explicit surfaces extracted, but only discrete surface probabilities. This paper describes the application of this representation to object recognition, based on Bayesian compositional hierarchies.
A multiframe 2D-to-3D video georegistration algorithm
Targeting from video relies upon precise image and video registration. Historically, the technology to automate this georegistration has operated using 2D transform spaces under the often naive assumption that the imaged geometry is planar. The author previously demonstrated a fast 2D-to-3D registration algorithm that removes this assumption, provided a digital elevation model (DEM) is available. Whereas the previous algorithm operated independently on each frame of a video sequence, a new 2D-to-3D algorithm is proposed that exploits the structural consistency of the imaged geometry across frames. This work presents this novel algorithm and explores its efficacy in reducing targeting error.
Advanced Concepts/Algorithms in ATR IV
icon_mobile_dropdown
Automatic tracking system with target classification
In this paper, we propose an overall target tracking scheme performing image stabilization, detection, tracking, and classification in the IR sensored image. Firstly, in the image stabilization stage, a captured image is stabilized from visible frame-to-frame jitters caused by camera shaking. After that, the background of the image is modeled as Gaussian. Based on the results of the background modeling, the difference image between a Gaussian background model and a current image is obtained, and regions with large differences are considered as targets. The block matching method is adopted as a tracker, which uses the image captured from the detected region as a template. During the tracking process, positions of the target are compensated by the Kalman filter. If the block matching tracker fails to track targets as they hide themselves behind obstacles, a coast tracking method is employed as a replacement. In the classification stage, key points are detected from the tracked image by using the scale-invariant feature transform (SIFT) and key descriptors are matched to those of pre-registered template images.
Advanced Concepts/Algorithms in ATR V
icon_mobile_dropdown
Component-based target recognition inspired by human vision
Yufeng Zheng, Kwabena Agyepong
In contrast with machine vision, human can recognize an object from complex background with great flexibility. For example, given the task of finding and circling all cars (no further information) in a picture, you may build a virtual image in mind from the task (or target) description before looking at the picture. Specifically, the virtual car image may be composed of the key components such as driver cabin and wheels. In this paper, we propose a component-based target recognition method by simulating the human recognition process. The component templates (equivalent to the virtual image in mind) of the target (car) are manually decomposed from the target feature image. Meanwhile, the edges of the testing image can be extracted by using a difference of Gaussian (DOG) model that simulates the spatiotemporal response in visual process. A phase correlation matching algorithm is then applied to match the templates with the testing edge image. If all key component templates are matched with the examining object, then this object is recognized as the target. Besides the recognition accuracy, we will also investigate if this method works with part targets (half cars). In our experiments, several natural pictures taken on streets were used to test the proposed method. The preliminary results show that the component-based recognition method is very promising.
Discrimination of classes of ships for aided recognition in a coastal environment
For naval operations in a coastal environment, detection of boats is not sufficient. When doing surveillance near a supposedly friendly coast, or self protection in a harbor, it is important to find the one object that means harm, among many others that do not. For this, it is necessary to obtain information on the many observed targets, which in this scenario are typically small vessels. Determining the exact type of ship is not enough to declare it a threat. However, in the whole process from (multi-sensor) detection to the decision to act, classification of a ship into a more general class is already of great help, when this information is combined with other data to assist an operator. We investigated several aspects of the use of electro-optical systems. As for classification, this paper concentrates on discriminating classes of small vessels with different electro-optical systems (visual and infrared) as part of the larger process involving an operator. It addresses both selection of features (based on shape and texture) and ways of using these in a system to assess threats. Results are presented on data recorded in coastal and harbor environments for several small targets.
Incremental learning in automatic target recognition
Chaitanya Raju, Karthik Mahesh Varadarajan, Aditya Kothari, et al.
ATR in two dimensional images is valuable for precision guidance, battlefield awareness and surveillance applications. Current ATR methods are largely data-driven and as a result, their recognition accuracy relies on the quality of training dataset. These methods fail to reliably recognize new target types and targets in new backgrounds and/or atmospheric conditions. Thus, there is a need for an ATR solution that can constantly update itself with information from new data samples (samples may belong to existing classes, background clutter or new target classes). In the paper, this problem is addressed in two steps: 1) Incremental learning with Fully Adaptive Approximate Nearest Neighbor Classifier (FAAN) - A novel data structure is designed to allow incremental learning in approximate nearest neighbor classifier. New data samples are assimilated at reduced complexity and memory without retraining on existing data samples, 2) Data Categorization using Data Effectiveness Measure (DEM) - DEM of a data sample is a degree to which each sample belongs to a local cluster of samples. During incremental learning, DEM is used to filter out redundant samples and outliers, thereby reducing computational complexity and avoiding data imbalance issues. The performance of FAAN is compared with proprietary Bagging-based Incremental Decision Tree (ABAFOR) implementation. Tests performed on Army ATR database with over 37,000 samples shows that while classification accuracy of FAAN is comparable to ABAFOR (both close to 95%), the process of incremental learning is significantly quicker.
Gauge features for curvilinear target recognition
Curvilinear targets are common in many imaging modalities. Detection of such targets can be challenging because of their multiscale structure, their frequent obscuration in natural imagery, their turns, intersections, and merges, and the prevalence of false positive detections based on local information. Using a spatial spectroscopy approach, we introduce image analysis methods that use the concept of gauge frames to simplify the identification of curvilinear targets. Fast computational approximation methods are described for gauge fields, and an experiment is described illustrating the power of higher-order derivatives for understanding even relatively simple geometric structures. Methods for extracting coherent curvilinear objects that exploit the larger-scale commonalities of points in the object are described.
Polarimetric and Infrared Processing for ATR
icon_mobile_dropdown
Polarimetric calibration and its influence on target recognition performance
Fully polarimetric radars that use polarization diversity on transmit and receive and thus provide the full scattering matrix, are subject to effects like cross-talk and channel imbalance. These distortions have to be eliminated by means of a polarimetric calibration in order to warrant compatibility between training data and testing data that were measured at different times or even by different radar sensors. It is shown for different types of classification features (geometric, statistical, polarimetric, structural) how an insufficient PolCal may influence the ATR performance.
Estimation and detection in degree of polarization images perturbed by detector noise and non uniform illumination
Active imaging systems that illuminate the scene with polarized light and acquire two images in two orthogonal polarizations yield information about the intensity contrast and the Orthogonal State Contrast (OSC) in the scene. However, in real systems, the illumination is often spatially or temporally non uniform. We first study the influence of this non uniformity on estimation performances. We derive the Cramer Rao Lower Bound and determine a profile likelihood-based estimator. We demonstrate the efficiency of this estimator and compare its performance with other standard estimators as a function of the degree of non-uniformity of the illumination. Concerning target detection, illumination non uniformity creates artificial intensity contrasts that can lead to false alarms. We derive the Generalized Likelihood Ratio Test (GLRT) detectors when intensity information is taken into account or not, and determine the relevant expressions of the contrast in these two situations. These results are used to determine in which cases taking intensity information in addition to polarimetric information is relevant or not.
A presentation of ATR processing chain validation procedure of IR terminal guidance version of the AASM modular air-to-ground weapon
D. Duclos, N. Quinquis, G. Broda, et al.
Developed by Sagem (SAFRAN Group), the AASM is a modular Air-To-Ground "Fire and Forget" weapon designed to be able to neutralise a large range of targets under all conditions. The AASM is composed of guidance and range enhancement kits that give bombs, already in service, new operational capabilities. AASM Guidance kit exists in two different versions. The IMU/GPS guidance version is able to achieve "ten-meter class" accuracy on target in all weather conditions. The IMU/GPS/IR guidance version is able to achieve "meter class" accuracy on target with poor precision geographic designation or in GPS-denied flight context, thanks to a IR sensor and a complex image processing chain. In this night/day IMU/GPS/IR version, the terminal guidance phase adjusts the missile navigation to the true target by matching the image viewed through the infrared sensor with a target model stored in the missile memory. This model will already have been drawn up on the ground using a mission planning system and, for example, a satellite image. This paper will present the main steps of the procedure applied to qualify the complete image processing chain of the AASM IMU/GPS/IR version, including open-loop validation of ATR algorithms on real and synthetic images, and closed-loop validation using AASM simulation reference model.
Development of a demeaning filter for small object detection in infrared images
Hwal-Suk Lee, Seokkwon Kim, Je Hee Lee, et al.
The demeaning filter detects a small object by removing a background with a mean filter as well as the covariance of an object and backgrounds. The factors considered in the design of the demeaning filter are the method of demeaning, which involves subtracting the local mean value from all pixel values, and the acquisition of templates for both the object and the background. This study compares the sliding window method and the grid method as a demeaning method, and studies the method of acquisition of an object template. Moreover, a method involving the use of previous frames, a mean filter, and an opening operation are studied in an effort to acquire a background template. Based on the results of this study, a practical design of a demeaning filter that is able to detect a small object in an IR image in real time is proposed. Experiment results demonstrate the superiority of the proposed design in detecting a small object following a 2-D Gaussian distribution even under severe zero-mean Gaussian noise.
Robust automatic target tracking based on a Bayesian ego-motion compensation framework for airborne FLIR imagery
Carlos R. del-Blanco, Fernando Jaureguizar, Narciso García, et al.
Automatic target tracking in airborne FLIR imagery is currently a challenge due to the camera ego-motion. This phenomenon distorts the spatio-temporal correlation of the video sequence, which dramatically reduces the tracking performance. Several works address this problem using ego-motion compensation strategies. They use a deterministic approach to compensate the camera motion assuming a specific model of geometric transformation. However, in real sequences a specific geometric transformation can not accurately describe the camera ego-motion for the whole sequence, and as consequence of this, the performance of the tracking stage can significantly decrease, even completely fail. The optimum transformation for each pair of consecutive frames depends on the relative depth of the elements that compose the scene, and their degree of texturization. In this work, a novel Particle Filter framework is proposed to efficiently manage several hypothesis of geometric transformations: Euclidean, affine, and projective. Each type of transformation is used to compute candidate locations of the object in the current frame. Then, each candidate is evaluated by the measurement model of the Particle Filter using the appearance information. This approach is able to adapt to different camera ego-motion conditions, and thus to satisfactorily perform the tracking. The proposed strategy has been tested on the AMCOM FLIR dataset, showing a high efficiency in the tracking of different types of targets in real working conditions.
Tracking of multiple objects under partial occlusion
Bing Han, Christopher Paulson, Taoran Lu, et al.
The goal of multiple object tracking is to find the trajectory of the target objects through a number of frames from an image sequence. Generally, multi-object tracking is a challenging problem due to illumination variation, object occlusion, abrupt object motion and camera motion. In this paper, we propose a multi-object tracking scheme based on a new weighted Kanade-Lucas-Tomasi (KLT) tracker. The original KLT tracking algorithm tracks global feature points instead of a target object, and the features can hardly be tracked through a long sequence because some features may easily get lost after multiple frames. Our tracking method consists of three steps: the first step is to detect moving objects; the second step is to track the features within the moving object mask, where we use a consistency weighted function; and the last step is to identify the trajectory of the object. With an appropriately chosen weighting function, we are able to identify the trajectories of moving objects with high accuracy. In addition, our scheme is able to handle partial object occlusion.