基于半监督生成对抗网络的毫米波雷达跨域手势识别

Cross-domain gesture recognition in millimeter-wave radar based on semi-supervised generative adversarial networks

基于深度学习的毫米波雷达手势识别以其免接触、保护隐私和环境依赖性低等特点受到越来越多的关注。但是目前的学习方法大都采用全监督方法,其性能受限于雷达数据的获取和标注,且其学习样本都来源于单一环境,极大的影响了不同场景下的迁移能力,因此本文提出了一种基于半监督生成对抗网络的跨域手势识别方法。首先,通过数据预处理,提取动态混合特征时间图(DFTM)以消除环境干扰,并且对手势动态特征进行更加全面的表征;其次,结合毫米波信号特点进行数据增强,进一步扩充数据量,提高模型泛化能力;第三,针对实际应用中可获得的标记数据通常较少的问题,提出并构建了一个改进半监督生成对抗网络,在原始GAN的基础上增加了分类器,通过生成数据帮助提高分类器辨别能力,同时利用源域中的少量标记数据和目标域中的大量未标记数据,实现域无关的手势识别。实验结果表明,对于新用户、新环境和新位置的平均手势识别准确率分别达到98.21%、95.23%和97.6%。与现有其他手势识别方法相比,本文所提方法在只有少量标记数据的情况下也能达到较高的跨域手势识别准确率,为后续毫米波雷达人机交互提供了新的研究思路。

Millimeter-wave radar gesture recognition based on deep learning has attracted more and more attention due to its characteristics of contact-free,privacy protection and low environmental dependence.However,most of the current learning methods use fully supervised methods,whose performance is limited by the acquisition and annotation of radar data,and their learning samples all come from a single environment,which greatly affects the transfer ability in different scenarios.Therefore,this paper proposes a A cross-domain gesture recognition method based on semi-supervised generative adversarial networks.First,through data preprocessing,the dynamic mixed time feature map(DFTM)is extracted to eliminate environmental interference and more comprehensively characterize the dynamic characteristics of gestures;secondly,data enhancement is performed based on the characteristics of millimeter wave signals to further expand the amount of data and improve the model.Generalization ability;thirdly,in order to solve the problem that the labeled data available in practical applications is usually less,an improved semi-supervised generative adversarial network is proposed and constructed.A classifier is added on the basis of the original GAN to help improve the performance by generating data.The discriminative ability of the classifier simultaneously utilizes a small amount of labeled data in the source domain and a large amount of unlabeled data in the target domain to achieve domain-independent gesture recognition.Experimental results show that the average gesture recognition accuracy for new users,new environments and new locations reaches 98.21%,95.23%and 97.6%respectively.Compared with other existing gesture recognition methods,the method proposed in this article can achieve high cross-domain gesture recognition accuracy even with only a small amount of labeled data,providing new research ideas for subsequent millimeter wave radar human-computer interaction.