| 摘要: |
| 目前基于毫米波雷达的人体动作识别方法存在人体动作特征利用不足导致识别精度低，与多维特征融合方式简单导致模型参数量增加的问题。针对以上问题，提出了一种基于毫米波雷达的三维特征自适应融合人体动作识别方法。首先，对雷达回波信号进行时频分析以获取人体动作距离、多普勒与角度特征，并在时域上拼接构建三维特征数据集。然后，设计了一种带有特征自适应融合器的三分支卷积神经网络，实现对三维特征数据集的高维抽象特征提取与多维特征的自适应融合。最后，通过活动分类器得到人体动作检测结果。实验结果表明，所提方法的人体动作识别平均准确率可达到96.41%，受试者工作特征曲线下面积（Area Under the Curve, AUC）值可达到0.984，优于单维特征识别方法与多种三维特征融合方法。 |
| 关键词: 毫米波雷达 动作识别 时频分析 特征融合 卷积神经网络 |
| DOI：DOI:10.3969/j.issn.1672-2337.2024.05.012 |
| 分类号:TN957.51 |
| 基金项目:国家自然科学基金资助项目（No.62261014, 62171146）；广西创新驱动发展专项（No. 桂科 AA21077008）；广西自然科学杰出青年基金项目（No.2021GXNSFFA220004）；广西科技基地和人才专项（No. 桂科 AD21220112）；广西研究生教育创新计划项目（No.YCBZ2023137） |
|
| Human Motion Recognition Algorithm with Three⁃Dimensional Feature Adaptive Fusion Based on Millimeter⁃Wave Radar |
|
LI Chenghui, JIANG Junzheng, ZHOU Fang
|
|
1. School of Information and Communication, Guilin University of Electronic Technology, Guilin 541004, China;2. Hangzhou Institute of Technology, Xidian University, Hangzhou 311231, China;3. School of Electronic Engineering, Xidian University, Xi’an 710071, China;4. College of Information Engineering, China Jiliang University, Hangzhou 310018, China
|
| Abstract: |
| At present, the human motion recognition methods using millimeter?wave radar exist the problems of low recognition accuracy caused by underutilization of human motion features and the increased number of model parameters due to the simple fusion method with multi?dimensional features. In view of the above problems, a three?dimensional feature adaptive fusion method for human motion recognition based on millimeter?wave radar is proposed. First, the time?frequency analysis of radar echo signals is carried out to obtain the features of human action distance, Doppler and angle, and the 3D feature dataset is constructed by accumulating them in the time domain. Second, a three?branch convolutional neural network with a feature adaptive fusion device is designed to realize the extraction of high?dimensional abstract features from 3D feature dataset and the effective fusion of multi?dimensional features. Finally, the human motion detection results are obtained by the activity classifier. The experimental results show that the average accuracy of human action recognition of the proposed method can reach 96.41%, and the area under the curve （AUC）value can reach 0.984, which is better than the single?dimensional feature recognition method and a variety of three?dimensional feature fusion methods. |
| Key words: millimeter⁃wave radar human motion recognition time⁃frequency analysis feature fusion convolutional neural network |
