Deep learning-driven UFIR filter for enhanced GNSS/PDR pedestrian localization

Abstract

Pedestrian Dead Reckoning (PDR), when integrated with a Global Navigation Satellite System (GNSS), promises enhanced reliability for pedestrian navigation. However, such an integration, typically based on the Kalman Filter (KF), faces challenges due to noise disturbances and model uncertainties. To address these issues, a robust integration framework was introduced, designed to bolster navigation precision. This framework updates the state by leveraging prior N measurements within a finite horizon window by utilizing historical data through an unbiased finite impulse response (UFIR) filter. Furthermore, to mitigate disturbances stemming from system uncertainties, deep learning techniques are employed instead of conventional model-driven methods to ascertain the generalized noise power gain (GNPG) of the UFIR filter. Extensive field experiments, including normal walking, walk to--run transitions, and scenarios with noise interference, were conducted to validate the efficacy of the approach. The outcomes revealed significant enhancements in both the robustness and accuracy of pedestrian navigation, surpassing existing techniques.