A motion-pattern learning and state-aware compensation method for GNSS-denied three-dimensional positioning
Abstract
Low-cost inertial and barometric sensing offers an autonomous route to three-dimensional (3D) positioning when global navigation satellite system (GNSS) signals are unavailable or unreliable. In long-duration pedestrian dead reckoning, however, displacement error, heading drift, altitude drift, and local trajectory-shape mismatch still accumulate and degrade the reconstructed trajectory. Many compensation methods also use fixed parameters or global corrections, which limits their adaptability to individual motion patterns and motion states. To address these limitations, we propose a motion-pattern learning and state-aware compensation method for calibration-assisted GNSS-denied 3D pedestrian positioning. First, synchronised dead-reckoning and real-time kinematic (RTK) reference trajectories are used to construct supervised calibration labels for step-length, heading, and altitude errors. The learned parameters are then fixed, and the compensated trajectory is reconstructed from inertial, barometric, and PDR-derived features without using RTK positions as online measurements. Second, a state-aware compensation strategy adjusts the correction strength under straight-motion, turning, and vertical-motion-related conditions. Finally, a constrained local residual correction method refines the compensated trajectory by correcting bounded short-window residual deformation. Two RTK-referenced field scenarios were analysed: a sports-field route for complete module-level evaluation and a non-closed-loop natural-slope route for stronger height variation. On the sports-field route, the proposed method achieved horizontal mean, RMSE, median, 75th-percentile, and 95th-percentile errors of 0.48, 0.64, 0.37, 0.59, and 1.11 m, respectively. The corresponding altitude absolute errors were 0.08, 0.09, 0.09, 0.11, and 0.15 m, and the 3D 75th-percentile error was 0.60 m. For the non-closed-loop natural-slope route, the 3D 75th-percentile error was 0.78 m. These results demonstrate comparable metre-level 3D accuracy on the tested calibration-assisted route scenarios.