| 摘要: |
| 针对传统雷达人体动作识别方法中特征提取能力不足和上下文建模困难的问题，提出了一种结合卷积神经网络(Convolutional Neural Network，CNN)和Swin Transformer的网络模型，用于有效识别分布式脉冲超宽带雷达数据中的人体动作。通过多分支的CNN对多个雷达的多个谱图、雷达数据的幅度和相位等特征进行提取和融合，利用Swin Transformer模块的多层自注意力机制对生成的特征映射进行上下文建模，提取具有高级语义信息的特征。采用代尔夫特理工大学（Technische Universiteit Delft）公开的数据集进行5折交叉验证，结果表明所提方法能够有效识别9类连续人体动作，识别准确率达到98.2%。 |
| 关键词: 分布式脉冲超宽带雷达 人体动作识别 卷积神经网络（CNN） Swin Transformer 特征融合 |
| DOI：10.20079/j.issn.1001-893x.230511004 |
|
| 基金项目:辽宁省兴辽英才计划项目(XLYC1907134)；辽宁省教育厅项目（LJKZ0174& LJKZ0173） |
|
| Human Activity Classification Based on Distributed Ultra瞱ideband Radar Combined with CNN-win Transformer |
| ZHANG Lili,JIA Dezhen,PAN Tianpeng,LIU Yanjuan |
| (College of Electronic Information Engineering,Shenyang Aerospace University,Shenyang 110136,China) |
| Abstract: |
| In response to the shortcomings of insufficient feature extraction capability and difficulties in contextual modeling in traditional radar-ased human activity classification methods,a network model is proposed,which combines convolutional neural network (CNN) and Swin Transformer to effectively recognize human activity in distributed pulse ultra-ideband radar data.Multiple branches of CNN are employed to extract and fuse features from multiple spectrograms of multiple radars,as well as the amplitude and phase of radar data.The multi-ayer self-ttention mechanism of the Swin Transformer module is utilized to model the generated feature maps in order to extract features with high-evel semantic information.The experimental results,obtained through 5-old cross-alidation on a publicly available dataset from Delft University of Technology,demonstrate that the proposed method can accurately identify nine categories of continuous human activity with a recognition accuracy of 98.2 |
| Key words: distributed pulse ultra-ideband radar human activity recognition convolutional neural network(CNN) Swin Transformer feature fusion |
