Publications

Assistive and Rehabilitation
Robotics Lab

Domestics

도메인 일반화를 적용한 웨어러블 센서 기반 피험자 독립 보행 모드 분류
Author
고동민, 이기욱
Journal
대한기계학회 춘계학술대회 바이오공학부문
Year
2025
Human sensor data collected during walking exhibits significant inter-individual variability, which often degrades the performance of deep learning models when classifying the locomotion environments of unseen subjects. To address this, we interpret such inter-subject variability as a domain shift and apply domain generalization techniques to mitigate its impact on classification accuracy. We investigated two widely adopted methods for addressing domain shift: Correlation Alignment (CORAL) and Domain Adversarial Neural Networks (DANN). Using signals from two thigh-mounted IMUs that measure hip joint angle and angular velocity, we classified four types of locomotion: level walking, ±5° incline/decline walking, and stair ascent/descent. Data was collected from five participants, and user-independent evaluations were conducted to assess model generalization to unseen individuals. A baseline convolutional neural network (CNN) achieved a classification accuracy of 96.58% ± 0.46% (mean ± SD across leave-one-group-out folds). Incorporating CORAL and DANN improved the accuracy to 96.96% ± 0.36% and 97.00% ± 0.83%, respectively, while a combined model using both techniques achieved 97.76% ± 0.87%. These findings demonstrate that domain generalization methods can effectively mitigate domain shift and enhance classification robustness, supporting their potential integration into wearable robotic systems such as exosuits for subject-independent locomotion mode classification.