Hand gesture recognition from sEMG signals based on similarity learning

Abstract

EMG (Electromyography) signals provide critical information on muscle activity and have long been used in prosthetics and healthcare. With the rise of wearable technologies, EMG is now widely applied in human-computer interaction, rehabilitation, and emerging remote control applications such as drone operation and gaming. The primary objective of this study was to accurately identify hand gestures utilizing limited user electromyography (EMG) data, thereby improving the applicability and extending the potential use cases of EMG-based systems in real-world scenarios. This study also evaluates the real-time capabilities of a deep learning–based approach for hand gesture recognition using limited user-specific EMG data. The proposed method makes use of deep learning methods, namely convolutional neural networks and transfer learning to learn the similarity between samples of EMG signals to predict hand gestures. A pre-trained network is fine-tuned through a progressive layer-freezing strategy to extract robust, subject-specific features. Subsequently, Support Vector Machines (SVM) are employed to map feature similarities, which inform a k-Nearest Neighbors (kNN) voting mechanism for final gesture classification. The findings indicate viability of the proposed method even when utilizing a limited dataset collected from the user. On the Myo Dataset, involving 17 subjects and 7 distinct hand gestures, the proposed approach outperformed the three approaches proposed in two related works as well as an approach based on a recurrent neural network. Further evaluation on the NinaPro DB5 (Exercise B) dataset demonstrated that the architecture consistently exceeds the classification accuracy of four state-of-the-art methods. Experimental results demonstrate that the proposed approach outperforms all compared methods in terms of classification accuracy. Finally, computational benchmarking demonstrates that the system's inference latency remains well within the established temporal thresholds for real-time applications cited in existing literature. Several promising directions for future research have been identified. These include real-time implementation and embedded deployment for practical prosthetic and robotic applications, systematic evaluation of a broader range of pretrained convolutional and transformer-based architectures within the multi-pass fine-tuning framework, and the incorporation of complementary sensor modalities — such as IMU sensors, depth cameras, and mechanomyography — through sensor fusion approaches to enhance classification robustness and reduce inter-subject variability.