AM-WiGeR：一种基于注意力机制增强残差网络的Wi-Fi手势识别方法

摘要：提出了一种基于注意力机制增强残差网络的Wi-Fi手势识别方法。该方法从Wi-Fi设备采集的信道状态信息（CSI）中提取多普勒频移（DFS）作为特征源，通过共轭矩阵相乘、天线对选择及相位偏移消除等方法，从中提取能有效表征手势信息的向量参数，并进一步通过短时傅里叶变换获得对应的DFS频谱图特征。然后，以DFS频谱图特征作为AM-RES模型的输入，进行人体手势识别。实验结果表明，相较于现有方法，所提AM-WiGeR方法对不同朝向手势识别的平均准确率达到96.22%，取得了良好的手势识别效果。

关键词：手势识别；CSI；DFS；注意力机制

Abstract: A Wi-Fi gesture recognition method based on Attention Mechanism Enhanced Residual Network (AM-RES) is proposed, termed AM-WiGeR (AM-RES based Wi-Fi Gesture Recognition). Channel state information (CSI) collected from widely deployed Wi-Fi devices is utilized, from which Doppler frequency shift (DFS) is extracted as the feature source. Through conjugate matrix multiplication, antenna pair selection, and phase offset elimination, vector parameters that can effectively characterize gesture information are extracted from the DFS, and the corresponding DFS spectrogram features are further obtained via short-time Fourier transform. The DFS spectrogram features are then fed into the AM-RES model for human gesture recognition. Experimental results demonstrate that, compared with existing methods, the proposed AM-WiGeR method achieves an average recognition accuracy of 96.22% for gestures performed in different orientations, indicating satisfactory gesture recognition performance.

Keywords: gesture recognition; CSI; DFS; attention mechanism