Comparative Study of ML, DL and Temporal Models in Plantar Gait Classification

This study classifies seven gait patterns (squatting, squat-to-stand, walking, going upstairs, going downstairs, standing, sitting) using plantar pressure time-series data. Following detailed data preprocessing (segmentation, window construction, feature engineering), we evaluated traditional machine learning (Random Forest, Gradient Boosting), fundamental deep learning (CNN, LSTM, Transformer, hybrid models), and advanced deep learning methods (contrastive learning, MiniRocket, ST-GCN variants). Results show: Gradient Boosting and ensemble models (ML) achieved 95.3% accuracy; hybrid models (DL) reached 94.9%; bilateral-fusion Transformer exceeded 95%; ST-GCN variants underperformed. Temporal feature-based deep learning methods demonstrated superior accuracy and generalization in gait classification, providing empirical support for model selection in gait recognition.