| 引用本文: | 丁晨旭, 张远辉, 孙哲涛, 刘 康. 基于FMCW雷达的人体复杂动作识别[J]. 雷达科学与技术, 2020, 18(6): 584-590.[点击复制] |
| DING Chenxu， ZHANG Yuanhui， SUN Zhetao， LIU Kang. Human Activity Classification Based on FMCW Radar[J]. Radar Science and Technology, 2020, 18(6): 584-590.[点击复制] |
|
| 摘要: |
| 针对利用摄像机进行人体动作识别时易受视距和光线影响等问题，提出一种基于FMCW雷达的人体复杂动作识别方案。首先基于FMCW信号模型对雷达采样数据采用一种以RDM（Range Doppler Map）向速度维投影的方式逐帧构建微多普勒谱图，继而基于微多普勒谱图来提取用于表征整个动作频谱相关信息的8种特征矢量。最后，基于雷达实测数据，以贝叶斯超参数调整优化后的支持向量机作为分类器，分析利用所提取的单特征矢量以及特征矢量组合来进行分类时对分类准确率的影响，用以筛选最优异的特征矢量组合。实验结果表明，从微多普勒谱图中所提取的特征矢量皆可直观地表述整个动作过程的特性，且利用最终筛选得到的最优异的特征矢量组合对已知个体和未知个体的9种动作进行识别，识别准确率分别高达99.07%和96.76%。 |
| 关键词: 毫米波雷达 人体动作识别 微多普勒 特征提取 支持向量机 |
| DOI：DOI:10.3969/j.issn.1672-2337.2020.06.002 |
| 分类号:TN958.94 |
| 基金项目:浙江省自然科学基金（No.LY19F010007） |
|
| Human Activity Classification Based on FMCW Radar |
|
DING Chenxu， ZHANG Yuanhui， SUN Zhetao， LIU Kang
|
|
China Jiliang University, Hangzhou 310018, China
|
| Abstract: |
| To solve the problems such as the influences of sight distance and light when using camera to perform human body motion recognition, a human activity classification approach based on FMCW radar is proposed. First, based on the FMCW signal model, the captured radar ADC raw data is projected to the velocity dimension by RDM (range Doppler map) to construct a frame-by-frame micro-Doppler spectrum. Then, the entire 8 kinds of feature vectors of motion spectrum related information is extracted based on the micro-Doppler spectrum. Finally, for the measured radar signal, an optimized support vector machine with Bayesian super parameter adjusted is used as the classifier to evaluate the classification accuracy of the extracted single feature vectors and their feature vector combination. The best combination set of feature vectors is selected for the human activity classification system. Experimental results show that the feature vectors extracted from the micro-Doppler spectrogram can intuitively express the characteristics of the entire motion, and the best feature vector combination set is used to classify the 9 kinds of known activities. Recognition rate of 9 activities of known and unknown individuals are 99.07% and 96.76% respectively. |
| Key words: millimeter wave radar human motion recognition micro-Doppler feature extraction support vector machine |
