基于度量学习的毫米波雷达少样本人体动作识别  

Metric Learning-Based Millimeter Wave Radar Human Action Recognition with Few Samples

作　　者：龚树凤[1] 施汉银 闫鑫悦 吴哲夫[1] GONG Shufeng;SHI Hanyin;YAN Xinyue;WU Zhefu(College of Information Engineering,Zhejiang University of Technology,Hangzhou Zhejiang 310012,China)

机构地区：[1]浙江工业大学信息工程学院,浙江杭州310012

出　　处：《传感技术学报》2024年第11期1921-1930,共10页Chinese Journal of Sensors and Actuators

基　　金：浙江省自然科学基金重点项目(LZ22F010005);浙江省教育厅科研项目(Y201839636)。

摘　　要：针对目前基于毫米波雷达的人体动作识别方法普遍需要大量的样本数据且计算复杂度较高的问题,提出了一种基于度量学习的毫米波雷达少样本人体动作识别方法。该方法首先对采集到的人体动作回波信号进行背景帧差处理得到校准后的帧数据,然后对其进行二维傅里叶变换(2D-FFT)获得距离-多普勒图,再对距离-多普勒图基于速度维投影法进行逐帧拼接来构造微多普勒时频谱图(DTM),最后使用基于残差的度量学习原型网络对8类人体动作的微多普勒时频谱图进行训练验证,实现了人体不同动作的识别。实验结果表明,所提方法在只有30个训练样本的情况下,8类动作的平均识别准确率可达到99.05%。To address the problem that current millimeter wave radar-based human action recognition method generally requires a large amount of sample data and has a high computational complexity,a metric learning-based millimeter wave radar human action recognition method with few samples is proposed.The method first performs background frame difference processing on the collected human action echo signals to obtain calibrated frame data,then performs two-dimensional Fourier transform(2D-FFT)on them to obtain distance-Doppler maps,then performs frame-by-frame stitching on the Range-Doppler maps based on velocity-dimensional projection to construct micro-Doppler time-frequency maps(DTM),and finally uses a residual-based metric-learning prototype network to construct micro-Doppler time-frequency maps for eight types of human.Eventually,a residual-based metric learning prototype network is used to train and validate the micro-Doppler time-frequency spectrograms for eight types of human actions.Experimental results show that with only 30 training samples,the average recognition accuracy of the proposed method can reach 99.05%for the eight types of actions.

关 键 词：毫米波雷达 人体动作识别 度量学习 背景帧差 

分 类 号：TN911.7[电子电信—通信与信息系统] TN957.51[电子电信—信息与通信工程]