Comparative Study of Machine Learning Methods for Lower-Limb Gait Recognition Based on The HuGaDB Dataset

Abstract

Lower-limb gait recognition plays an important role in rehabilitation robotics, intelligent healthcare, human motion analysis, and wearable devices. However, comparative studies of different machine learning models under the same public dataset and unified experimental conditions remain limited. To address this issue, this study employs the Human Gait Database (HuGaDB) public multimodal dataset, using accelerometers, gyroscopes, and electromyography (EMG) sensors as input features to compare the performance of three classification models: Support Vector Machine (SVM), Long Short-Term Memory (LSTM), and Random Forest (RF). Data preprocessing, feature extraction, and training–testing dataset partitioning were performed prior to model training. The three models were trained and evaluated under identical experimental conditions. Model performance was assessed using Confusion Matrix Recall F1-score Accuracy, Precision. The experimental results show that the RF model achieved the best overall performance. RF provides the highest classification accuracy and stability under the experimental conditions of this study, making it more suitable for lower-limb gait recognition based on multimodal data collected from IMUs and EMG sensors. The results also demonstrate the effectiveness of multimodal sensor fusion for gait recognition and provide useful reference for the selection of machine learning models in intelligent rehabilitation systems, wearable devices, and other human activity recognition applications.