| Over the past few decades, multi-sensor data fusion has been applied to a broad range of problems in many different areas including object detection and recognition, target tracking, remote sensing, medical diagnosis, robotics, and autonomous vehicles. Researchers have recognized that the synergistic combination of data from multiple sensors can provide a more robust and complete view of the object of interest than can be achieved by a single-sensor system. Further advances require a better understanding of the science behind different multi-sensor data fusion systems. This dissertation presents our research on several selected issues concerning multi-sensor data fusion systems that have recently received significant attention. In particular, we focus on the novel signal processing design and performance evaluation techniques for three popular systems: the multi-sensor image fusion system, the multi-input and multi-output (MIMO) radar system and the distributed sensor network.;Quantitatively measuring the performance of a multi-sensor image fusion system is a complicated but important task. We focus on the theoretical analysis of three correlation-based fused image quality measures (FIQMs) when they are used to judge the performance of weighted averaging image fusion algorithms. Our analysis shows that when we change the power of the desired signal or the noise in the source images, these correlation-based FIQMs exhibit some undesired behaviors. In addition, we develop a novel statistic to score the effectiveness of FIQMs for the detection task in light of practical measurements from human perception experiments. The performance of the proposed monotonic test is demonstrated via Monte Carlo simulations. We also show the application of the proposed method to evaluate potential FIQMs in a specific target detection experiment.;The second part of this dissertation considers the joint location and velocity estimation problem in a multi-target non-coherent MIMO radar system. The Cramer-Rao bound (CRB) is a useful tool for evaluating the performance of radar systems, as it provides the mean square error lower bound for any unbiased estimation. In this dissertation, we focus on a multi-target case, in which a non-coherent MIMO radar system is considered. This case has not yet been studied by others. We investigate the joint location and velocity estimation of multiple targets, and the Cramer-Rao bound for a two-target case is derived and evaluated. This bound gives us theoretically achievable joint estimation performance for a sufficient number of antennas.;The third part of the dissertation considers the design of change detection methods using observations from distributed sensor networks, where each node has access to local observations and is only allowed to communicate with its neighbors. Results apply to monitoring large systems like the electrical grid but they also apply generally to cases where sensors monitor changes in a random field. Using our algorithms, all the nodes will reach a consensus on the test in the end. First, we study the distributed change detection problem for distributions that can be represented as Gaussian graphical models. We propose two distributed tests. The first distributed test is a natural approach which simply applies the generalized likelihood ratio test (GLRT) to smaller size local clusters in the graph. The second method employs the pseudo-likelihood as a surrogate function for the global likelihood. Next, we consider the fault detection problem for measurements following the errors-in-variables (EIV) model. The standard approach to parameter estimation in such problems is known as total least squares (TLS). Recently, a competing approach known as total maximum likelihood (TML) was proposed and was shown to provide promising performance gains in various estimation problems. Following these works, we derive the TLS based GLRT and the TML based GLRT, which are specifically tailored for the smart grid structure. |
