Model Fitting Filter

In this paper, we present a new nonlinear filter based on variational inference, which we have named the model fitting filter (MFF). The core of MFF is to represent or fit the nonlinear measurement likelihood probability with a linear Gaussian probability (LGP) characterized by some parameters that need to be turned. Further, MFF is analytically established on LGP, which is a little similar to the simple linear system filter. Specifically, the nonlinear filtering performance is well improved by utilizing these turned parameters to correct the state estimation. At the same time, MFF does not require the square root decomposition of covariance matrix, so it avoids the inborn defect of numerical instability, which exists more or less in UKF or DDF. The superior performance of MFF is numerically tested by a target tracking problem with nonlinear models.