MyoLite: Compressible Neural Networks Based on Biokinematics and Self-attention for Fingers Motion Tracking

With more attention drawn to brain-machine interface (BCI) technology, sEMG-based human-machine interaction (HMI) is used more often in virtual reality (VR), prostheses, exoskeletons, robots, and medical rehabilitation. However, users are no longer satisfied with the recognition of a few discrete gestures, but tend to vividly and continuously track the finger movements for getting better interaction experience. In this paper, with a sparse sEMG armband, we propose a compressible neural network, MyoLite, which is an explainable end-to-end network model including electrode decomposition, myo encoding and myo decoding, to achieve continuous motion tracking of fingers by bioanatomical analysis. Besides, we analysed the linkage relationship of finger movements through self-attention mechanism to compress and generalize the network model. With a public dataset, MyoLite obtains a higher accuracy of finger joint motion tracking than previous methods. The results show the root-mean-square error (RMSE) reaches to 6.10° and the fit is approaching 90.93%. Furthermore, while maintaining the accuracy, the bionic analysis-based compression strategy achieves 30% reduction in the amount of weighting parameters and 31.9% reduction in the floating-point operations per second (FLOPs).