A comprehensive evaluation of machine learning and deep learning models for stair-climbing wheelchair activity recognition

Abstract

Human Activity Recognition (HAR) is vital for enhancing assistive technologies such as stair-climbing wheelchairs, which cater to individuals with mobility challenges. This study investigates optimal machine learning and deep learning models for classifying human activity related to stair-climbing wheelchairs, which are essential for enhancing mobility in assistive technologies. A dataset of 5,872 samples across 18 sensor-derived features was preprocessed using normalization, one-hot encoding, and SMOTE to address class imbalance. Eight models—MLP, CNN, LSTM, BiLSTM, Transformer, CatBoost, LightGBM, and TabNet—were trained and evaluated using an 80/20 train-test split. CatBoost and LightGBM achieved the highest accuracy (99.83%) with inference times of 8 ms and 7 ms respectively. Deep learning models such as MLP and CNN also performed well, while the Transformer exhibited poor compatibility with the dataset. Machine learning models, especially CatBoost and LightGBM, demonstrated both high accuracy and computational efficiency, making them suitable for real-time applications in assistive technologies. This work provides essential insights for implementing efficient HAR systems in mobility-assistive devices and can inform future designs of autonomous wheelchair platforms.