LipSense:基于CSI相位差的自适应唇语识别方法

LipSense:Adaptive Lip Motion Recognition Method Based on CSI Phase Difference

针对目前基于CSI幅度的行为识别方法对细粒度动作不敏感及传统的动作分割算法存在动作起止点判断不准确的问题,提出了一种基于相位差的自适应唇语识别方法。首先,利用MIMO系统的空间分集,构造出对细粒度动作更敏感的相位差信号,解决了细粒度动作感知不易的问题;然后利用OFDM子载波的频率分集来补偿小尺度衰落效应,以增强信号的空间感知能力;再次,通过研究唇语信号独特的变化模式,提出基于迭代方差曲线的自适应唇语分割模型;最后设计用于表示和匹配唇语特征的识别技术。在真实环境下的实验表明,所提方法在单人情况下平均识别精度达到94.3%,三人的综合识别准确率可达85.7%。在实验者佩戴金属饰品、网络被他人共享、以及周围有人移动等对比实验下,系统仍能实现较高的准确率。

To address the problem that the current behavior recognition method based on CSI amplitude is insensitive to fine-grained lip motions and the traditional motion segmentation algorithms have inaccurate judgments of action start and end points,an adaptive lip motion recognition method is proposed based on CSI phase difference.Firstly,the spatial diversity of the MIMO system is used to construct the phase difference signal that is more sensitive to fine-grained motions,solving the problem that finerained motions are difficult to detect.Then the frequency diversity of OFDM sub-carriers is used to compensate the small-scale fading effect to enhance the spatial sensing ability of signals.Thirdly,by investigating the unique variation pattern of the lip signal,an adaptive lip motion segmentation model based on iterative variance curves is proposed.Finally,a recognition technology for representing and matching lip features is designed.Experiments in real environment show that the average recognition accuracy of the proposed method is 94.3%in the case of single person and 85.7%in the case of three people.The system can still achieve high accuracy under the comparative experiments of the subject wearing metal ornaments,the network being shared by others,and people in the neibouring area moving around.