Distributed detection and fusion in parallel sensor architectures

Parallel distributed detection system consists of several separate sensor-detector nodes (separated spatially or by their principles of operation), each with some processing capabilities. These local sensor-detectors send some information on an observed phenomenon to a centrally located Data Fusion Center for aggregation and decision making. Often, the local sensors use electro-mechanical, optical or RF modalities and are known as "hard'' sensors. For such data sources, the sensor observations have structure and often some tractable statistical distributions which help in weighing their contribution to an integrated global decision. In a distributed detection environment, we often also have "humans in the loop.''. Humans provide their subjective opinions on these phenomena. These opinions are labeled "soft'' data. It is of interest to integrate "soft'' decisions, mostly assessments provided by humans, with data from the "hard" sensors, in order to improve global decision reliability. Several techniques were developed to combine data from traditional hard sensors, and a body of work was also created about integration of "soft'' data. However relatively little work was done on combining hard and soft data and decisions in an integrated environment. Our work investigates both "hard'' and "hard/soft'' fusion schemes, and proposes data integration architectures to facilitate heterogeneous sensor data fusion. In the context of "hard'' fusion, one of the contributions of this thesis is an algorithm that provides a globally optimum solution for local detector (hard sensor) design that satisfies a Neyman-Pearson criterion (maximal probability of detection under a fixed upper bound on the global false alarm rate) at the fusion center. Furthermore, the thesis also delves into application of distributed detection techniques in both parallel and sequential frameworks. Specifically, we apply parallel detection and fusion schemes to the problem of real time computer user authentication and sequential Kalman filtering for real time hypoxia detection. In the context of "hard/soft'' fusion, we propose a new Dempster-Shafer evidence theory based approach to facilitate heterogeneous sensor data fusion. Application of the framework to a number of simulated example scenarios showcases the wide range of applicability of the developed approach. We also propose and develop a hierarchical evidence tree based architecture for representing nested human opinions. The proposed framework is versatile enough to deal with both hard and soft source data using the evidence theory framework, it can handle uncertainty as well as data aggregation.