Robust centralized multi-sensor fusion using cubature information filter

Noise statistics is crucial to the estimation fusion problem for the sensor network. In this paper, a centralized multi-sensor fusion for the network with non-Gaussian measurement noises is considered. The heavy-tailed Student-t distribution is chosen to model the non-Gaussian measurement noise for each sensor. By using a hierarchical model for the Student-t distribution, Wishart distribution and Gamma distribution are introduced to model the uncertainties of the noise parameters. The recursive update equations for the combined information state, combined information matrix, and noise parameters are derived by using variational Bayesian approach together with spherical cubature integration handling nonlinearities. Specifically, the information state and information matrix are obtained based on the fused local estimates of non-Gaussian noise parameters from all the sensors. A multi-sensor target tracking example illustrates the robustness of the proposed multi-sensor fusion method to the outliers.