IFNSUD 2013 - Special Session on Information Fusion in Networked Systems under Unknown Dependencies

Standard Bayesian fusion techniques rely on the assumption that estimates are independent. However, this is often violated. For example, in multisensor systems, common prior information and correlated measurement noises across sensors have to be dealt with and, in distributed tracking algorithms for same targets, common process noise is an important problem. On the one hand, keeping track of the correlation structure between estimates is often not possible. Especially in a decentralized multisensor network, estimates are processed locally and are fused with other local estimates. Since the individual sensor nodes are not aware of the global network topology, double counting can cause strong correlations between local estimates to be fused. On the other hand, correlations often entail high computational costs. For example, standard Kalman filter methods in simultaneous localization and mapping (SLAM) applications are hardly tractable, when non-sparse cross-covariance matrices of thousands of landmarks have to be managed. Decoupling and decorrelating partial estimates can then reduce the complexity significantly. Dealing with unknown correlations is still very challenging. Existing methods are often highly conservative or suboptimal. Also, a generalization to nonlinear estimation remains an open question, where even uncorrelated states can be stochastically dependent. New ideas, improvements of existing approaches, and new solutions to this highly relevant problem can be presented and discussed within the Special Session Information Fusion in Networked Systems under Unknown Dependencies.