| Summary: | Gestures are an important way to conduct human-computer interaction. The key problem of gesture recognition depending on sEMG (surface electromyography) is how to achieve high recognition accuracy when there are many types of gestures to classify. To solve this problem, first, two basic models were constructed. One is the ConvEMG model based on dense connectivity, the Inception module and depthwise separable convolution; and the other is the LSTMEMG model based on a bidirectional LSTM (Long Short-Term Memory). Then, the basic models were improved with a multistream fusion strategy which utilizes the correlation between gestures and muscles and the complementary advantages of models. To facilitate comparison with others’ models, the models proposed in this paper were tested on the public dataset NinaPro DB5, and the improved model named MultiConvEMG achieves an accuracy of 92.83% for 41 gestures, which is superior to its counterparts in the literature on the same dataset. In addition, experiments containing signal acquisition and gesture recognition were carried out for further testing and evaluation. Experimental results show that all models can achieve an accuracy of more than 95% for 31 gestures, and these models have their own strengths in accuracy, immediacy or training cost. All models built in the paper support using sEMG for end-to-end recognition, which means that artificial features are not needed in the processes and data augmentation or IMU devices are not relied on. In other words, our models outperform and have lower application costs than many known models.
|
|---|
