MULTI-TASK LEARNING-BASED PEDESTRIAN IDENTIFICATION AND GAIT STATE CLASSIFICATION FOR NORMAL AND MUSCLE-FATIGUED GAITS

초록

Gait patterns exhibit subject-specific biometric characteristics and can be used for subject identification. However, physiological condition changes, such as muscle fatigue, may alter gait characteristics in real-world environments. In this paper, we propose a hard-parameter-sharing-based multi-task learning (MTL) framework to perform simultaneous subject identification and normal/fatigued gait classification utilizing inertial measurement unit (IMU)-based gait signals. The proposed model learns local patterns and temporal dependencies from 30-axis IMU time series signals through a shared backbone consisting of Conv1D and long short-term memory layers and simultaneously performs 12-class closed-set subject identification and binary normal/fatigued gait classification through two task-specific output branches. In experiments with 12 healthy adult male participants, the proposed model achieved an accuracy of 99.96% and a macro-F1-score of 99.96% for subject identification across five-fold subject-dependent validations. For normal/fatigued gait classification, the model achieved a balanced accuracy of 99.93% and an AUC of 99.99%. In a leave-one-subject-out (LOSO) subject-independent validation, which evaluated generalization performance for unseen subjects, the model achieved a balanced accuracy of 64.17% and an AUC of 69.83% for normal/fatigued gait classification. These results demonstrate that the proposed MTL framework can simultaneously perform subject identification and normal/fatigued gait classification utilizing IMU-based gait signals.

키워드