An improved particle filter and its application to an INS/GPS integrated navigation system in a serious noisy scenario

Abstract

For loosely coupled INS/GPS integrated navigation systems with low-cost and low-accuracy microelectromechanical device inertial sensors, in order to obtain enough accuracy, a full-state nonlinear dynamic model rather than a linearized error model is much more preferable. Particle filters are particularly for nonlinear and non-Gaussian situations, but typical bootstrap particle filters (BPFs) and some improved particle filters (IPFs) such as auxiliary particle filters (APFs) and Gaussian particle filters (GPFs) cannot solve the mismatch between the importance function and the likelihood function very well. The predicted particles propagated through inertial navigation equations cannot be scattered with certainty within the effective range of current observation when there are large drift errors of the inertial sensors. Therefore, the current observation cannot play the correction role well and these particle filters are invalid to some extent. The proposed IPF firstly estimates the corresponding state bias errors according to the current observation and then corrects the bias errors of the predicted particles before determining the weights and resampling the particles. Simulations and practical experiments both show that the proposed IPF can effectively solve the mismatch between the importance function and the likelihood function of a BPF and compensate the accumulated errors of INSs very well. It has great robustness in a serious noisy scenario.