Fall-Related Gait Recognition Using Surface Electromyography with Transfer Learning

Recent advances in gait analysis have highlighted the challenges posed by individual differences, making it difficult to apply consistent methods across diverse individuals. This study presents a system for decoding simulated gait patterns using surface electromyography (EMG), addressing issues of generalization and the heavy training demands of neural networks for recognizing gait patterns in new individuals. Data from seven healthy participants were collected, including normal and two fall-related gaits, with myoelectric signals recorded from eight lower limb muscles. To enhance model performance, transfer learning (TF) was utilized, and eight adaptation strategies were tested, each with varying target training sizes. In cross-subject evaluations, even with few target samples, the proposed method showed strong generalization and lower training demands. This research emphasizes the effectiveness of combining TF with neural networks for reliable gait recognition across individuals, providing valuable insights for improving the accuracy and efficiency of EMG-based gait analysis systems.