| 摘要: |
| 针对微多普勒特征识别人体动作的局限性，基于调频连续波（Frequency Modulated Continuous Wave，FMCW）雷达采用深度学习方法对人体动作识别，提出了一种特征融合卷积神经网络结构。利用FMCW雷达采样的人体动作回波数据分别构建出时间-距离特征和微多普勒特征图，将这两种特征图作为输入数据分别经由输入层进入卷积层，经Batch Normalization层、ReLU激活函数和最大池化层计算之后完成特征降维，然后对两种降维后的特征进行融合，融合后的特征图再经过卷积层和池化层计算获得更深层次的特征，最后经过两个全连接层，在输出层完成人体动作识别。采用英国格拉斯哥大学公开的数据集进行10折交叉验证，实验结果显示，与单一特征域的识别准确率相比，采用两种特征融合的结构进行人体动作识别的准确率提升了1%，验证了该模型的有效性。 |
| 关键词: 人体动作识别 调频连续波雷达 特征融合卷积神经网络 时间-距离特征 微多普勒特征 |
| DOI： |
|
| 基金项目:国家自然科学基金资助项目（61671310）；航空科学基金（2019ZC054004）；辽宁省兴辽英才计划项目基金(XLYC1907134)；辽宁省百千万人才工程项目基金(2018B21)；辽宁省教育厅项目（LJKZ0174） |
|
| Human activity recognition with FMCW radar based on fusion feature convolutional neural network |
| ZHANG Lili,LIU Bo,QU Lele,LIU Yuxuan |
| (College of Electronics Information Engineering,Shenyang Aerospace University,Shenyang 110136,China) |
| Abstract: |
| For the limitation of human activity recognition based on micro-Doppler feature，a fusion feature convolutional neural network(CNN) is proposed using a frequency modulated continuous wave(FMCW) radar and deep learning method.The time-range feature and micro-Doppler feature diagrams are constructed by using the echo data from FMCW radar.The extracted features are used as the input data to enter the convolutional layer respectively.Through the Batch Normalization layer,the ReLU activation function and the maximum pooling layer，the feature reduction is completed.Then the two reduced features are fused.The fused feature diagram passes through the convolutional layer and the pooling layer to obtain deeper features,and finally passes through the 2 fully connected layers to complete the human action recognition in the output layer.A 10-fold cross-validation is performed based on the University of Glasgow “Radar Signatures of Human Activities” open dataset and the result shows the recognition accuracy of the proposed structure is improved by 1% when compared with that of a single feature，which proves the validity of the model. |
| Key words: human activity recognition frequency modulated continuous wave radar fusion feature convolutional neural network time-range feature micro-Doppler feature |
