Model Distribution in Decentralized Multi-Sensor Data Fusion
Timothy Berg, Hugh F Durrant-Whyte · 1991
This paper considers the problem of data fusion in a decentralized and distributed network of multi-sensor processing nodes. A decentralized and distributed Kalman filter is formulated. This filter needs no central processor; globally optimum estimates are obtained at each mode by requiring an underlying centrality of the state space rather than a centralized topology. Model distribution reduces the computational burden at each node by allowing each node a local model that is most appropriate to the dynamics of its observations. The combination of model distribution and decentralization yields a robust and efficient parallel processing network. The problem of communicating and assimilating relevant estimates directly between nodes with different models and state subspaces is solved using internodal transformations. The same intermodal transformations are used to relate models directly between nodes. Ralationships between the transformations and network configuration considerations are discussed.