Abstract: In order to improve the accuracy of gesture recognition algorithm using sEMG ( Surface Electromyography) signals and solve the problem of accuracy affected by various features extracted, a sEMG's recognition method based on MB-CNN (Multi-Branch Convolutional Neural Networks) is proposed. Firstly, a MYO armband is utilized to sample sEMG signals of 8 different gestures. Secondly, the sliding window method is used to detect active segment of sEMG signals and the original training samples with the size of 64×8 are obtained. Thirdly, as a comparative experiment, seven different time-domain and frequency-domain features are extracted from original samples and machine learning algorithms are used to achieve the gesture recognition. Finally, in the case of avoiding conventional feature extraction, a MB-CNN model is constructed to achieve the gesture recognition and the accuracy of test set gains 97.89% . Experiment shows the proposed method is efficient and feasible for gesture recognition.

Key words: surface electromyography ( sEMG), multi-branch convolutional neural networks ( MB-CNN), gesture recognition, MYO armband