| 摘要: |
| 针对人体动作识别微多普勒雷达数据量有限的问题，本文提出基于梯度惩罚的沃瑟斯坦生成对抗网络(WGAN-GP)进行雷达数据增强，实现深度卷积神经网络(DCNN)在样本数量较少时可以得到有效训练。首先对人体各种动作的线性调频连续波雷达回波数据进行预处理得到微多普勒时频谱图像，然后采用WGAN-GP进行时频谱图像数据增强，最后利用生成的图像对DCNN进行训练。实验结果表明使用WGAN-GP可以有效解决雷达数据不足的问题，从而提高DCNN人体动作识别准确率。 |
| 关键词: 人体动作识别 微多普勒 沃瑟斯坦生成对抗网络 深度卷积神经网络 |
| DOI：DOI:10.3969/j.issn.1672-2337.2022.02.011 |
| 分类号:TN958 |
| 基金项目:国家自然科学基金（No.61671310）； 航空科学基金（No.2019ZC054004）； 辽宁省兴辽英才计划项目(No.XLYC1907134)； 辽宁省百千万人才工程项目(No.2018B21) |
|
| Human Activity Recognition Based on WGAN-GP in Micro-Doppler Radar |
|
QU Lele, WANG Yutong
|
|
College of Electronic Information Engineering, Shenyang Aerospace University, Shenyang 110136, China
|
| Abstract: |
| To solve the problem of limited number of micro-Doppler radar data samples, Wasserstein generative adversarial network-GP (WGAN-GP) is proposed to augment radar data for the human activity recognition, which can make the deep convolutional neural network (DCNN) to be trained effectively even with small number of data samples . Firstly, the micro-Doppler spectrogram images of various human activities are obtained by preprocessing the echo signal of linear frequency modulated continuous wave radar. And then WGAN-GP is adopted for the augmentation of micro-Doppler spectrogram images. Finally, the synthetic images are used to train the DCNN. Experimental results show that WGAN-GP can effectively solve the problem of insufficient radar data and improve the classification accuracy of human activity for DCNN. |
| Key words: human activity recognition micro-Doppler WGAN-GP DCNN |
