| 摘 要: 针对传统的基于FMCW雷达的人体动作识别算法复杂度较高,占用内存大的问题,提出了一种基于轻量化模型的雷达人体动作识别方法。将采集的雷达人体动作数据进行预处理得到微多普勒图,将其输入以多尺度特征融合模块改进的FasterNet模型,从不同角度出发完成图像特征的学习。将注意力机制引入FasterNet Block中,增强重要特征信息的提取,提高模型识别精度。实验结果表明,该方法动作识别准确率高达98.83%,且与传统的雷达人体动作识别方法相比具有更低的复杂度。 |
| 关键词: 毫米波雷达 微多普勒图 多尺度特征融合 注意力机制 部分卷积 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 福建省科技重大专项(GY-Z14012);福建理工大学科研启动基金项目(GY-Z21064,GY-Z21065) |
|
| A Lightweight Model-Based Radar Human Motion Recognition Method |
|
LIN Chengyan1, ZHOU Yujie1, NI Qinwei1, YUAN Yusheng1, ZHANG Lin1, CAI Zhiming1,2
|
(1.School of Electronic, Electrical Engineering and Physics, Fujian University of Technology, Fuzhou 350118, China;
2.National Demonstration Center for Experimental Electronic Information and ElectricalTechnology Education, Fujian University of Technology, Fuzhou 350118, China)
2221905076@smail.fjut.edu.cn; 352989926@qq.com; 2221905029@smail.fjut.edu.cn; 2221908016@smail.fjut.edu.cn; 9160481@qq.com; caizm@fjut.edu.cn
|
| Abstract: To address the issues of high complexity and large memory usage in traditional FMCW radar-based human motion recognition algorithms, this paper proposes a lightweight model-based radar human motion recognition method. First, the collected radar human motion data is preprocessed to obtain micro-Doppler spectrograms. Next, the micro-Doppler spectrograms are input into an improved FasterNet model with a mult-i scale feature fusion module to learn image features from different perspectives. Finally, an attention mechanism is introduced into the FasterNet Block to enhance the extraction of important feature information and improve the model’s recognition accuracy. Experimental results show that this method achieves a motion recognition accuracy of up to 98.83% , and compared to traditional rada-r based human motion recognition methods, it has lower complexity. |
| Keywords: millimete-r wave radar Micro-Doppler spectrogram mult-i scale feature fusion attention mechanism partial convolution |
