Decentralized Detection with Correlated Observations

We study the problem of decentralized detection when the observations are dependent conditioned on the hypothesis. The optimal rule for fusing the quantized versions of the observations received from the individual sensors requires complete statistical information about the quantized observations. Since such a fusion rule may not be easy to implement in practice, we propose a suboptimal detector that optimizes a deflection metric in the class of linear-quadratic detectors. Using simulations it is shown that when the observations are correlated the deflection-optimal detector outperforms the detector that ignores the correlation information completely.