The binary parallel decentralized detection system is the most popular architecture studied so far in the literature and the one investigated in the present study. It comprises a bank of local decision makers and a central information processor called the data fusion center. Each of the local decision-makers gathers observations about a binary hypothesis and makes a binary {H₀, H₁} decision which is transmitted over a noiseless communication channel to the data fusion center. At the data fusion center, the local decisions are combined into a final decision about the hypothesis, to satisfy a global performance index such as a Bayes risk or a Neyman-Pearson criterion.

In spite of the growing body of literature about decision fusion, several theoretical obstacles hinder the use of distributed detection architectures in practice. First, many proposed designs are too expensive computationally to be practically useful. Then, many engineering constraints and practical communication and synchronization requirements have not yet been considered. Our laboratory attempts to advance the understanding and resolution of some of these issues.

Among the topics that we have studied are the design of distributed detection architectures by employing optimal local decision makers and an optimal data fusion center independantly. this modification reduces the performance of the system when compared to a globally optimal design, but increases computational feasibility. We have also studied asynchronism in the communication links between the local detectors and the data fusion center and examined the effect of simple inexpensive decision memory on the decision rules and on the performance of the parallel distributed-detection system. Most recently we have been working on adaptive decision fusion and on decision fusion for mobile robot navigation.

Members Working on Distributed Detection

Joe Prosser, Xiaoxun Zhu