重心動揺を用いた個人識別における識別精度向上と特徴量分析

We have improved personal identification accuracy using dynamic center of pressure (COP) data during stepping on and stepping off the balance board. Wii Balance Board data was resampled at 50Hz, log-transformed, smoothed (Savitzky–Golay), and classified into five segments (S1–S5) based on X-axis acceleration. To reduce variability, only stabilization (S2), standing (S3), and pre-fall (S4) were analyzed. In 10-subject experiments using 36 features (including mean, standard deviation, and skewness of position and velocity), SVM (>99%) and random forest (>98%) outperformed XGBoost (~95%). SHAP analysis confirmed lateral dynamic characteristics in S2 and S4 contribute to classification, demonstrating high accuracy and interpretability.