A High Accuracy and Real-Time sEMG-based Hand Gesture Classifier using LDA-based Template Matching with Adaptive Majority Vote and Online Data Augmentation

Surface electromyography (sEMG)-based gesture classification is promising in many emerging applications such as prosthetic control and Augmented Reality (AR) / Virtual Reality (VR). While many existing sEMG-based hand gesture classifiers achieve high intra-session accuracy, they show very low inter-session and inter-subject accuracy which are more important in practical use. Another issue of the existing works is the usage of deep neural networks increases the accuracy but also brings up the complexity, making real-time classification difficult especially on embedded systems. In this work, an sEMG-based hand gesture classifier with high inter-session, inter-subject accuracy and real-time performance is proposed. A linear discriminant analysis (LDA)-based template matching method is proposed to improve the inter-session and inter-subject accuracy while maintaining low complexity. An adaptive window-size majority vote method and a calibration-oriented online data augmentation method are proposed to further improve the accuracy. The proposed classifier has been implemented on both GPU and FPGA-based embedded system demonstrating real-time performance. Compared to the state-of-the-art works, the proposed classifier achieves higher inter-session and inter-subject accuracy of 78.0% and 78.9% on Ninapro DB6 (95.2% and 96.6% on CapgMyo DB-b) with shorter classification time of 6.5ms on Ninapro DB6 (4.9ms on CapgMyo DB-b).