Neural Network-based Architectures for Target Surveillance and Tracking
This project, supported by grants from the Office of Naval Research (ONR) and the Ballistic Missile Development Organization (BMDO), is primarily aimed at designing neural network-based schemes for target surveillance and tracking. The principal objectives of the project are to (i) demonstrate the ability of trained neural networks to fuse information from diverse sensors to assist in simple implementations of target surveillance and tracking algorithms, and (ii) develop tracking architectures utilizing neural networks that are capable of tracking targets executing complex evasive maneuvers in clutter and noise environments. The major accomplishments and research advances being made in this project include the following:
demonstration of the feasibility of neural network-based approaches for tracking target maneuvers;
demonstration of the feasibility of multi-sensor data fusion for target tracking without attendant increases in computational complexity;
development of target tracking architectures that integrate a trained neural network with a Kalman filter to yield superior tracking performance without increased implementation complexity;
development of a new target motion model, termed "Equivalent Velocity Tracking Model", that facilitates neural network training for maneuver compensation and reliable tracking;
development of a new training scheme for multilayer neural networks based on simplex optimization and progressive network growing;
demonstration of the tracking performance in diverse target maneuver scenarios which is superior to performance delivered by existing schemes.
These results are established through rigorous mathematical analysis and extensive simulation experiments. In addition, an application of the neural network-based tracking approach to the recently developed benchmark problem for the beam steering of a phased-array radar system is also being made. Descriptions of the various studies conducted and the major results obtained are published in various reports and papers.
Image Restoration and Super-resolution
This project, supported by grants from the Air Force Office of Scientific Research (AFOSR), Air Force Wright Laboratory Armament Directorate, and the Army Research Laboratories (ARL), is primarily aimed at the design and development of novel image restoration and super-resolution algorithms that could be used for enhancing the resolution of various types of acquired imagery data. A particular emphasis is being placed on tailoring algorithms that are ideally suited for processing Passive Millimeter-Wave (PMMW) and Synthetic Aperture Radar (SAR) images. Images acquired from practical sensing operations typically suffer from poor spatial resolution due to the finite size of antenna or lens that makes up the imaging system and the consequent imposition of the underlying diffraction limits. Hence, some form of post-detection image processing directed to enhancing the resolution will often be needed before the obtained measurements can be used reliably for any decision-making objectives. Specific tasks that are undertaken in support of this goal include: (i) A detailed analysis of resolution challenges addressing several important questions such as how to quantify resolution in an image (acquired or processed), what effects will digital representation of data and selection of sampling rates will have on the overall resolution achievable, etc., (ii) development of systematic digital processing algorithms obtained using statistical optimization approaches and convex set theoretic methods for achieving resolution enhancement and super-resolution, and (iii) performance evaluation studies that include results of processing simulated image data and results of processing real PMMW and SAR image data. The major accomplishments and research advances made in this project include the following:
Since satisfactory restoration and super-resolution of different types of images can be achieved only through an intelligent tailoring of iterative processing algorithms, an analysis of some fundamental issues that need to be taken into account in synthesizing algorithms that afford simple digital implementation has been conducted and several useful conclusions have been developed.
Complete descriptions of several powerful restoration algorithms that aim at maximizing carefully chosen statistical performance measures (likelihood and posterior density) and that employ Projection Onto Convex sets (POCS) formulations have been obtained. Some modifications of these algorithms for handling blind deconvolution problem scenarios and for optimized implementations through a progressive upsampling scheme are also obtained.
Demonstration of resolution enhancements in several SAR and PMMW images obtained from state-of-the-art sensor platforms has been made. In addition, a quantitative evaluation of the resolution gains achieved from digital processing of the acquired data has also been developed.
These results have been obtained through rigorous mathematical analysis and extensive simulation experiments, and are published in various reports and papers.
Design of Data and Sensor Fusion Architectures
This project, supported by grants from the Office of Naval Research (ONR) and the Ballistic Missile Development Organization (BMDO), is aimed at the design of systematic architectures and algorithms for fusing dissimilar data collected from diverse sensors. Multisensor operations in several practical applications demand implementation of efficient fusion architectures that support intelligent integrated processing of incoming data streams which comprise of large volumes of data arriving at typically high rates without excessive computational complexities. Several fundamental studies directed to understanding the role of information contained in measured data and any a priori knowledge one may have about the scene, alternate representations of this information in dynamically varying environments, effects of including signal processing methods for modifying the sensor outputs, and methods of utilizing information representations in the design of fusion architectures are being investigated. A specific focus is given to the design of fusion architectures using neural networks and for facilitating improved target surveillance and guidance schemes.
Mobility Management for Wireless Networks
Several issues have become important due to recent developments in communication devices and services, especially with increase in wireless and wireline devices and networking options. These are congestion in area codes, number portability, need to support intelligent networking, wireless-wireline integration, and demand for new user services. For meeting the requirements needed to overcome associated problems, this project is aimed at the design of novel network management schemes that facilitate associating multiple devices with a person by assigning globally unique portable numbers. A primary focus is placed on studying a variety of possible database architectures with different levels of replication for implementing such person-based numbering schemes. Methods of queueing theory are used to evaluate the performance of these mobility management schemes.
Intelligent Management and Control Architectures for Broadband Networks
The evolution of service from basic long distance calling between two people to multimedia information exchange among people and machines has revolutionaized the topics of network control and network management. It is increasingly being recognized that an integrated approach to network management and control is needed to develop efficient architectures and protocols. In very broad terms, network control relates to procedures aimed at realizing optimal performance or to ensure a certain quality of service under dynamically changing traffic load and network conditions, and these procedures include adaptive routing, flow control, buffer and bandwidth management, access control, and congestion avoidance schemes. Network management, on the other hand, relates to all of the functions performed to manage network resources and has the following functional categories: (i) fault management (fault detection, trouble reports), (ii) performance management (performance monitoring, alerts), (iii) configuration management (network topology database, routing changes, bandwidth allocations), (iv) accounting management (traffic usage statistics, billing reports), and (v) security management (secured access, intrusion detection and recovery, authentication, encryption and secure message transfer). The principal goals in this project are to design advanced management and control architectures for networks with broadband service capabilities. An integrated approach to handling the management and control problems by employing intelligent system techniques such as neural networks, fuzzy logic, and expert system-based approaches. Both analytical modeling and simulation-based methods are employed to evaluate real-time performance quantities of interest and to validate the management and control support system under various multimedia traffic and service management scenarios.
Trajectory Generation and Classification Using Dynamic Neural Networks
Neural network techniques for static problems, such as functional mapping and pattern recognition, have experienced a great development during the past two decades. On the other hand, development of neural network techniques for dynamic problems, such as trajectory mapping and classification, have not seen such an advancement due to the fact that these tasks typically require dynamic neural nets ( networks with recurrent and feedback connections) which pose tremendous training complexities. In this project, we are examining various methods for training dynamic neural networks for learning spatio-temporal patterns. In particular, the trajectory generation problem that involves training a neural network to learn and autonomously replicate a specified time-varying periodic motion when started from arbitrary initial states is given a particular focus due to its varied applications in speech and control problems. We are also developing neural network-based trajectory classification schemes for recognition of trained spatio-temporal behavior in the face of noise and distortions.
Neural Network-based Control and Guidance Schemes
In this project, supported by grants from Rockwell International and Hughes Missile Development Corporation (HMDC), we are examining various neural network architectures (feedforward networks, recurrent networks, radial basis functions, support vector machines, etc.) for designing intelligent control and guidance schemes. Development of efficient training schemes that include error backpropagation and teacher forcing concepts are given a particular focus. Also being investigated are learning methods that do not require evaluation of error gradients and hence offer significant computational benefits. Applications to control of robots and industrial manipulators, power system control problems, and missile guidance schemes are being considered under this project.