Robust PCA-based Walking Direction Estimation via Stable Principal Component Pursuit for Pedestrian Dead Reckoning

This paper proposes an outlier-robust pedestrian walking direction estimation method. The outliers caused by the unexpected behavior of pedestrians, e.g., wiping sweat, are detected by exploiting the eigenvalue characteristics of the principal components obtained by principal component analysis (PCA) on the distribution of the acceleration measurements. Once the outliers are detected, we solve a robust PCA problem with a newly defined stable principal component pursuit (SPCP) in the inertial sensor measurement domain to recover a low-rank matrix from the acceleration measurement matrix. Eventually, from this outlier-removed low-rank matrix, we estimate the correct walking direction. The performance of the proposed method was evaluated through experiments on several behaviors defined as unexpected behavior of pedestrians. In the outlier-inducing scenario, the proposed method using robust PCA via SPCP outperformed the existing methods by about 70% with a mean error of 4.83°. Furthermore, in the extended scenario, the robust PCA via SPCP outperformed the existing methods by 70–78% with a mean error of 3.50°, improving the robustness of the PCA-based walking direction estimation method.