A MACHINE LEARNING APPROACH FOR STRIDE LENGTH ESTIMATION IN PEDESTRIAN DEAD RECKONING SYSTEMS

Abstract

Stride length estimation is a critical component in pedestrian dead reckoning (PDR) systems, as estimation errors accumulate and degrade positioning accuracy, particularly in GNSS-denied environments. This study proposes a machine learning model based on optimized Extreme Gradient Boosting (XGBoost) for accurate stride length estimation. The model employs a compact model configuration with shallow decision trees, regularization mechanisms, and a mean absolute error–oriented objective function to enhance robustness across varying gait patterns. Experimental evaluation on a public dataset shows that the proposed model achieves a mean absolute error (MAE) of 0.042 m, with a stable error distribution across different stride lengths. The proposed approach reduces stride length estimation errors by approximately 4–26%, depending on the reference method, under the same dataset and experimental conditions. These results demonstrate that the proposed model provides a practical, accurate, and computationally efficient solution for applications requiring precise stride length estimation, including PDR, gait analysis, indoor localization, and wearable health monitoring systems.