摘要
目的表面肌电信号可以直接反映用户的动作意图,近年来已经成为手势识别等人机交互任务的主要控制信号。然而,个体差异性使得用户模型不能通用,限制了其应用与发展,为此,本文提出一种新的跨个体对抗适应网络(cross-subject adversarial adaptation network,CAAN)。方法该网络包括特征编码器、手势分类器和个体分类器3个子模块,使用了新的对抗性适应训练方法训练网络,达到分离出个体私有特征的目标。CAAN网络在采集的数据集上进行训练和测试,数据集包括11名受试者的6种手势。结果本方法的手势识别准确率达到88.08%,通过比较,该方法的性能优于现有的方法。结论本文提出的CAAN网络可有效进行跨个体手势识别,为人机交互提供可靠的技术。
Objective Intra-subject hand gesture recognition based on surface electromyography has been extensively studied in current years.However,the gesture recognition on cross-subject tasks has more broad application prospects.So we propose a novel cross-subject adversarial adaptation network(CAAN),which fulfills the cross-subject gesture recognition task.Methods An adversarial adaptation training method is developed to train the network to encourage the emergence of the features that are discriminative and subject-independent.The CAAN is evaluated on the collected dataset(including six gestures from eleven subjects).Results The proposed method outperforms other methods,which achieving offline accuracies as 88.08%.Conclusions The proposed CAAN can effectively carry out cross-subject gesture recognition and provide reliable technology for human-computer interaction.
作者
朱九英
米红林
付佳杰
ZHU Jiuying;MI Honglin;FU Jiajie(Shanghai Technical Institute of Electronics&Information,Shanghai 201411)
出处
《北京生物医学工程》
2023年第2期124-129,共6页
Beijing Biomedical Engineering
基金
上海电子信息职业技术学院专项项目(Z22105)资助。
关键词
表面肌电信号
手势识别
人机交互
跨个体对抗适应网络
特征
surface electromyography
hand gesture recognition
human-computer interaction
cross-subject adversarial adaptation network
feature
