Unscented-Transformation-Based Distributed Nonlinear State Estimation: Algorithm, Analysis, and Experiments

The problem of fully distributed state estimation using networked local sensors is studied in this paper. Specifically, the scenario with general nonlinear process model and local sensing models is considered by extending the distributed hybrid information fusion (DHIF) algorithm proposed by Wang and Ren. Different from the extended Kalman filter-based approaches which require the computation of the Jacobian matrix at every time instant, the unscented transformation (UT) approach is adopted for such an extension to better characterize the statistics after nonlinear transformations. The extended algorithm inherits the advantages of the original DHIF algorithm for requiring only one communication iteration between every two consecutive time instants and for requiring no global information. As well recognized that the stability analysis in the distributed UT-based framework is challenging, in the special case where the sensing models are linear, it is also analytically shown that the local estimate errors are bounded in the mean square sense. Simulations are extensively studied to show the performance of the extended algorithm. More importantly, the effectiveness of the algorithm is also verified using real data collected in a robot tracking task with networked Vicon cameras.