Causal Bayesian Networks for Robust and Efficient Fusion of Information Obtained from Sensors and Humans

This paper discusses techniques for fusion in contemporary situation assessment applications. Such applications often require reasoning about phenomena that cannot be observed directly, but information about their effects (i.e. symptoms) can be accessed through the existing sensory and communication infrastructure. Reasoning about hidden phenomena requires interpretation of relevant observations. Observations can be of heterogeneous types and can originate from humans as well as various sensory systems. Interpretation in such settings can be very challenging, as there might exist complex dependences between different phenomena. In addition, we are often confronted with significant modeling and observation uncertainties. Particularly challenging is the fact that a large portion of such information often originates from humans. Consequently, it can be very difficult to obtain perception models that precisely describe the distributions of hidden phenomena and human reports. In this paper we show that Bayesian networks (BNs) are suitable for the development of fusion systems in such settings, because they can efficiently describe the monitoring domains. Moreover, BNs support construction of efficient and robust distributed fusion systems.