基于注意力残差网络的特征肌电图手势识别

Study in gesture recognition of feature electromyography based on the attention residual network

基于表面肌电信号的手势识别方法效果欠佳,提出一种信号的特征样本构造方法和注意力残差网络(ARN),用于识别NinaproDB1中的23类功能性抓握手势.样本构造过程中利用巴特沃斯滤波器对表面肌电信号进行一阶1 Hz低通滤波,去除噪声干扰并保留通频带和过渡带信号;用时间窗截取滤波信号生成表面肌电图;计算信号平均绝对值、方差和波长特征,融合生成特征肌电图.引入注意力机制和残差连接改进卷积神经网络,搭建ARN用于手势识别.将特征肌电图样本输入ARN,在测试集中的手势识别准确率为86.87%,证明构造的信号特征样本结合ARN识别实用手势的有效性.

The poor effect of current practical gesture recognition methods based on surface electromyographic signals,a signal feature sample construction method and attention residual network(ARN)were proposed to identify 23 types of functional grasp gestures in NinaproDB1.In the course of sample construction,the surface electromyography signal was low-pass filtered by the Butterworth filter to remove noise interference and meanwhile retain the signals of pass band and transition band.The filtering signal was truncated to generate surface electromyography by using the time window.The mean absolute value,variance and wavelength features of signal were simultaneously calculated and fused to generate the feature electromyogram.After feeding the feature electromyogram samples into ARN,the accuracy of gesture recognition on the test set reached 86.87%,proving the effectiveness in the constructed sample of signal feature combined with ARN to recognize practical gestures.