In recent years, machine learning has achieved significant advancements in inertial positioning, exceeding the accuracy of traditional methods based on Newtonian mechanics. This paper introduces a novel network architecture that integrates frequency-domain and time-domain information to improve trajectory prediction accuracy. The proposed approach leverages the periodicity and sparsity of frequency-domain processing to effectively model the long-time correlation in IMU data. Additionally, by introducing the sLSTM with a scalar update mechanism, the model leverages its enhanced time-series modeling capabilities to further reduce inertial positioning errors. Experimental evaluations on multiple public datasets validate the effectiveness of the proposed method. Specifically, on the RoNIN dataset, the absolute trajectory error (ATE) and relative trajectory error (RTE) are reduced by 43.0% and 13.1%, respectively.
To protect the results of the paper, we will release the core code after the paper is officially accepted. We appreciate your understanding.