基于双流特征融合的FMCW雷达人体连续动作识别

Continuous Human Motion Recognition Methodfor FMCW Radar Based on Two-Stream Feature Fusion

本文提出了一种基于双流特征融合的FMCW雷达人体连续动作识别方法。首先对人体动作雷达回波信号进行预处理得到距离时间域图与微多普勒时频谱图,之后分别对两个不同维度的图像进行主成分分析提取对应特征并选取相同时间段的主成分分析结果进行融合得到双流融合特征,最后将双流融合特征输入到Bi-LSTM网络中训练与测试,网络对每个时间段的输入特征产生与之对应的动作类别输出从而实现连续人体动作识别。实验结果表明,当采用双流融合特征作为Bi-LSTM网络的输入时平均识别准确率要高于只采用距离时间特征或微多普勒特征作为网络输入时的平均识别准确率。

A continuous human motion recognition method for FMCW radar based on two-stream fusion feature is proposed in this paper.Firstly,the range-time image and the micro-Doppler spectrogram image are obtained by preprocessing the radar echo data.And then the two-stream fused features are obtained from range-time image and micro-Doppler spectrogram image by the principal component analysis(PCA)method and the extracted features are fused in the same period.Finally,the two-stream fused features are put into the Bi-LSTM network for training and testing.And the network provides the corresponding action category and implements the continuous human motion recognition.The experimental results show that the average recognition accuracy of the proposed method using two-stream fused features as the input of Bi-LSTM network is higher than that of the network only using range-time feature or the micro-Doppler feature.