Abstract: Oriented to the high-performing myoelectric control systems,a surface Electromyography (sEMG) based gesture recognition approach based on dual-stream convolutional neural network was proposed.The discrete wavelet transform coefficients together with raw sEMG signals were input into two branches of a dual-stream convolutional neural network respectively for high-level feature learning.The high-level features learned by two branches were fused together via a high-level feature fusion module.The proposed framework was evaluated on three large-scale benchmark databases containing sEMG signals collected from 50 to 52 hand gestures.Experimental results showed that the majority voting accuracy of the proposed framework reached 97.9%,81.3%,and 82.4% on three benchmark databases respectively when recognizing all hand gestures.Moreover,the gesture recognition accuracy of the proposed framework based on sliding windows significantly outperformed the state-of-the-art deep neural networks.

Key words: gesture recognition, dual-stream convolutional neural network, discrete wavelet transform, surface electromyography, myoelectric control systems

摘要： 面向高性能的肌电控制系统,提出一种基于双流卷积神经网络的肌电信号手势识别方法,其从原始表面肌电信号中提取离散小波变换系数,与原始表面肌电信号分别作为双流卷积神经网络两个分支的输入进行高层特征学习,最终通过一个高层特征融合模块对两个分支学习得到的高层特征进行融合。所提方法在3个包含50～52类手势动作表面肌电信号的大规模基准数据集中,识别所有手势动作的投票准确率分别达到97.9%,81.3%,82.4%,且在3个数据集中基于不同长度滑动采样窗口的手势识别准确率均显著超越了近年来本领域相关研究工作所提出的深度神经网络模型。

关键词: 手势识别, 双流卷积神经网络, 离散小波变换, 表面肌电信号, 肌电控制系统