基于FMCW雷达的多通道特征融合人体动作识别方法

针对采用单一特征进行人体动作识别准确率不高的问题，提出了一种基于调频连续波(Frequency Modulated Continuous Wave, FMCW)雷达的多通道特征融合人体动作识别方法。通过对FMCW雷达回波数据进行预处理，得到人体动作的距离参数与多普勒参数，构建出距离-时间特征谱图和多普勒-时间特征谱图数据集。为了进行人体动作特征的充分提取与精确识别，改进了单通道输入的传统卷积神经网络结构，把部分残差连接结构和跨阶段部分连接结构进行了优化应用至雷达人体动作识别领域，设计出端到端的CSP-FCNN(Cross Stage Partial-Fusion Convolutional Neural Network)多通道融合卷积神经网络。采用公开数据集进行实验，结果表明所提方法有效解决了单一特征动作识别信息量欠缺以及网络提取特征不充分的问题，识别准确率较单一特征识别方法提高了5%以上。

Multi-channel feature fusion for human activity recognition based on FMCW radar

For the problem of low accuracy of human activity recognition(HAR) using a single feature,a method for HAR based on multi-channel feature fusion of frequency modulated continuous wave(FMCW) radar is proposed.The range and Doppler parameters of human activity are obtained by FMCW radar echo data pre-processing,and the range-time and Doppler-time spectra datasets are constructed.The convolutional neural network(CNN) structure is improved to fully extract and accurately identify human activity features.The end-to-end Cross Stage Partial-Fusion Convolutional Neural Network(CSP-FCNN) multi-channel fused CNN is designed by applying partial residual connection structure and cross-stage partial connection structure in the field of radar human activity recognition.Experiment using open datasets shows that the proposed method effectively solves the problem of insufficient information for single feature motion recognition and insufficient network extracting features,and the recognition accuracy is improved by more than 5% compared with that of other single feature recognition methods.