Enhanced Gesture Recognition via Interpretable Temporal Convolutional Networks Utilizing Spatio-Temporal sEMG Features

• Published in: 2024 IEEE Conference on Pervasive and Intelligent Computing (PICom) pp. 61 - 66
• Main Authors: Ao, Xiaohu, She, Jinhua, Wang, Feng, Kawata, Seiichi, Fukushima, Edwardo Fumihiko, Woo, Jinseok
• Format: Conference Proceeding
• Language: English
• Published: IEEE 05.11.2024
• Subjects: Accuracy; Computational modeling; Convolutional neural networks; Explainable Deep Learning; Gesture recognition; Intent recognition; Predictive models; Real-time systems; sEMG; Spatiotemporal phenomena; Stability analysis; TCN; Time complexity
• DOI: 10.1109/PICom64201.2024.00015

Rehabilitation robots and human-robot interaction systems are receiving increasing attention in modern healthcare and engineering. Motion intent recognition techniques based on surface electromyography (sEMG) signals provide important support for these systems. However, motion intent recognition models rarely focus on the spatiotemporal properties of sEMG signals simultaneously. There is a lack of interpretable analysis of these spatiotemporal properties. The purpose of this study is to perform gesture recognition of sEMG signals from the Ninapro database using a temporal convolutional network (TCN). The sEMG signals are converted into sEMG images, which are used as input to the TCN model. The TCN model is utilized to perform gesture recognition on these images. The results show that after fusing the spatial and temporal features of the sEMG signal, the recognition accuracy of the TCN model is significantly higher than that of models using only spatial features or only temporal features. To understand the decision-making process of the model, we used an interpretable method based on Shapley values to perform a detailed interpretive analysis of the recognition results. The experimental results show that the TCN model has significant accuracy and stability in the sEMG signal gesture recognition task, while the Shapley value-based interpretable method can effectively reveal the contribution of different features to the prediction results, thus improving the transparency and user trust in the model. This study not only provides a new technical tool for sEMG signal gesture recognition but also demonstrates the value of interpretable methods in deep learning models.