Robust multi-object sensor fusion with unknown correlations

Distribution and decentralisation of fusion operations are key to network centric operations (NCOs) and distributed data fusion algorithms (DDF) have been developed to support them. These algorithms fuse data collected locally with state estimates propagated from other nodes. If the full advantages of NCOs are to be realised, these algorithms should exploit local information only: no single node, for example, should be an oracle which must maintain the entire state of the network. Uhlmann argued that many of these could be overcome if suboptimal solutions were used and proposed a principled suboptimal algorithm known as Covariance Intersection (CI). CI has proved to be a very powerful and general method for fusing data in arbitrary networks and has been used in a range of distributed and other applications where full correlation structures cannot be maintained. However, CI only utilizes the mean and covariance of the estimates and cannot exploit any additional distribution information such as the number of modes. The generalisation of CI to general probability distributions was first proposed by Mahler and independently derived by Hurley. We investigate the generalisation Covariance Intersection for multi-object posteriors by considering specific forms of multi-object posterior and their first-order moment densities, Probability Hypothesis Densities, as a prerequisite study for determining tractable implementations.