Distributed Detection and Data Fusion

The initial paper on the subject of distributed detection, by Tenney and Sandell, showed that under a fixed fusion rule, for two sensors with one bit outputs, the optimal Bayes sensor decision rule is a likelihood ratio test. It has been shown that the optimal fusion rule for N sensors is a likelihood ratio test on the data received from the sensors. Reibman and Nolte and Hoballah and Varshney have generalized the results to N sensors with optimal fusion, again with the restriction of one bit sensor outputs; this has been relaxed later to multi-bit quantizations. In this “primer” talk we explore a number of issues in distributed detection, including some pathologies, the benefits of fusion, optimal design, structures for decision flow, consensus, sensor biases, feedback, deliberate obfuscation (i.e., security) and censoring. We also devote some time to distributed estimation (i.e., fusion for tracking): why is it difficult and what seems to work best?