Abstract
This study classified walking environments (hallway, grass, sandy beach, uphill slope, downhill slope) using an insole-type gait sensor and conducted a fundamental investigation of walking environment estimation technology with an eye toward applications in the tourism sector. Diverse walking environments exist in tourist destinations, parks, and historic streetscapes, where mobility load and safety directly impact the comfort and accessibility of the tourist experience. This study constructed walking environment classification models using five methods: Random Forest, SVM, XGBoost, 1D-CNN, and Transformer Encoder, and evaluated them via cross-validation among subjects. The results showed that the deep learning models 1D-CNN and Transformer Encoder demonstrated relatively high classification accuracy, maintaining strong generalization performance even for unseen subjects. Furthermore, through misclassification tendency analysis and visualization using t-SNE, we clarified differences in feature distributions and challenges in environmental identification. These findings suggest potential applications in barrier-free evaluations of tourist destinations, fatigue risk estimation, and route recommendation/mobility support within smart tourism.
