Efficient neural inertial localization

Abstract Recently, human localization and navigation have drawn increasing attention, with a surge of data-driven methods for long-term pedestrian position estimation and steadily improving results. However, their heavy computation limits widespread use, especially on mobile devices. To address this, we present efficient neural inertial localization (ENILoc) for pedestrian position estimation from inertial measurement units. ENILoc combines the simplicity and speed of linear models with strong accuracy, while operating without external infrastructure and thus avoiding common privacy concerns. Unlike many existing approaches, ENILoc contains no Transformer blocks or convolutional layers; instead, it is built from a small number of linear layers. We validate ENILoc through extensive comparisons on four public datasets. Results show that ENILoc outperforms state-of-the-art alternatives, trains 2.5–6$\times $ faster, and reduces memory usage by 17–31$\times $. In addition, inference latency is reduced by up to 8$\times $, enabling practical on-device deployment. Finally, localization accuracy improves substantially, with significant gains in both absolute trajectory error and relative trajectory error, confirming ENILoc’s effectiveness for long-term indoor positioning.