| 摘要: |
| 传统深度学习依赖大量训练数据，而雷达数据的收集耗时费力，且难以涵盖所有手势类别，针对上述问题，本文提出一种基于Transformer元学习网络的毫米波雷达手势识别方法。首先对毫米波雷达采集的手势回波进行预处理得到多普勒时频图(DTFM)，将DTFM输入到特征提取模块，该模块结合卷积神经网络局部特征提取和Transformer全局上下文建模能力，能够准确提取手势动作对应的DTFM特征。然后构建基于Transformer和序列池化的非线性分类器，执行特征级比较以衡量不同手势样本之间的相似度。最后将比较结果送入到分类模块中进行手势类别预测。实验结果表明所提方法在小样本情况下可以有效地提高毫米波雷达手势动作识别准确率。 |
| 关键词: 手势识别 毫米波雷达 Transformer网络 元学习 |
| DOI： |
| 分类号:TN958.95 |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Millimeter-wave Radar Gesture Recognition Method Based on Transformer Meta-learning Network |
|
|
|
|
| Abstract: |
| The traditional deep learning-based radar gesture recognition methods require a large amount of training data. However, collecting radar data is time-consuming and difficult to encompass all gesture categories. To address this challenge, this paper proposes a millimeter-wave radar gesture recognition method based on a Transformer meta-learning network. First, the gesture echoes collected by millimeter-wave radar are preprocessed to generate Doppler time-frequency maps (DTFM). The DTFM are then fed into a feature extraction module, which combines the local feature extraction capability of the convolutional neural network (CNN) with the global context modeling strength of the Transformer, enabling it to accurately extract DTFM features of different gesture actions. Additionally, a nonlinear classifier based on Transformer and sequence pooling is designed to perform feature-level comparisons, measuring the similarity between different gesture samples. Finally, the comparison results are fed into the classification module to predict and output the final gesture categories. The experimental results indicate that the proposed method can effectively enhance the gesture recognition accuracy under the condition of limited samples. |
| Key words: gesture recognition millimeter wave radar transformer network meta learning |