Date of Award

2016

Degree Type

Dissertation

Degree Name

Doctor of Philosophy in Electrical Engineering

Department

Electrical, Computer, and Biomedical Engineering

First Advisor

Steven M. Kay

Abstract

The detection problem of a target based on radio-frequency (RF) and infrared (IR) data sources is addressed in this dissertation. The target is assumed to radiate an RF signal to multiple widely distributed sensors in space and is concurrently imaged using multiple frames of an IR sensor whose field-of-view is fixed with respect to the background. The observations contain additive noise in both sources due to the physical nature of the problem and sensor imperfections.

First, we define accurate signal models to be used in hypothesis testing for the RF and IR data sources. Second, the generalized likelihood ratio test (GLRT) statistics for both RF and IR detection problem are derived. Once the GLRT statistics are found, the optimal integration of test statistics is done to obtain the integrated GLRT detector. Although the integration is done optimally, the integrated GLRT requires; the joint maximization of a highly nonlinear statistic over target motion parameters, which is computationally expensive, and the submission of all RF data observed at local sensors to the central processor (CP), which requires high communication bandwidth and a CP having large memory. The random basis functions (RBF) approach is proposed as a computationally efficient method to reduce the computational complexity and data compression techniques are proposed for the distributed detection of the RF signal.

The RBF approach is applied to IR image sequences and it is shown that the reduction in computations is substantial as the dimension of the unknown parameter space is high. However, it causes some performance loss when compared to GLRT detector, thus, the approach requires higher signal-to-noise ratio to operate. Another possible use of the RBF approach is to apply it to reduce down the search space for the RF detector instead of applying it until convergence. Then, the GLRT for the RF data is run over the reduced down search space to obtain the multimodal estimates of the target motion parameters. This alternative representation of the multimodal detector is also implemented and the localization performance is tested. It is seen that the multimodal detector is more robust to partial occlusion when compared to detection using RF sensors only.

For the RF signal detection problem, the GLRT requires the submission of all data obtained at local sensors to the CP where the maximization takes place. Although this classical centralized detector has asymptotically optimal detection performance, submission of all observed data to the CP is practically infeasible as the RF detection problem is typically based on large data records. Therefore, distributed detection methods are considered for this problem and novel approaches based on Taylor expansions are proposed. Observed data is compressed into local test statistics at each sensor and transmitted to the CP for the formation of overall test statistic. The GLRT detector performance is used as the upper bound to assess the performance of the proposed compression techniques in this work.

Share

COinS