Data-driven inertial positioning method based on dual-end feature extraction

Inertial Navigation Systems (INS) can serve as an important alternative in environments where satellite-based positioning is limited because of their autonomy and independence from external signals. However, standalone INS based on low-cost inertial sensors is prone to cumulative drift because of limited sensor accuracy and complex noise characteristics. Most existing deep learning methods mainly preprocess front-end inertial input sequences while paying limited attention to the temporal structure of reference position labels, requiring the network to implicitly learn position variations at different time scales from a single supervision target. To address this issue, this paper proposes a data-driven inertial positioning method based on dual-end feature extraction. At the input end, variance-guided screening retains the original sequences of low-fluctuation channels, while Seasonal-Trend decomposition using Loess (STL) is applied to high-fluctuation channels. At the label end, each ground-truth position coordinate is decomposed into trend, seasonal, and residual components to construct structured supervision targets. A soft attention-enhanced Bidirectional Long Short-Term Memory network is then employed to predict the position components, and the final position is reconstructed through inverse normalization and component summation. Experiments are conducted on the Oxford Inertial Odometry Dataset (OxIOD) under five scenarios: Pocket, Trolley, Handheld, Running, and Slow Walking. The results show that the proposed method achieves lower position estimation errors than the selected recurrent-network baselines, with root mean square error values of 0.178 and 0.284 in the Pocket and Trolley scenarios, respectively. Additional experiments under different device carrying conditions and motion intensities further demonstrate the effectiveness of the dual-end feature extraction strategy under the evaluated experimental conditions.