【摘要】 雷达动作特征谱图对走、摔等较为宽幅的人体动作表征效果差，单一特征和不匹配特征数据结构的分类方法会降低动作识别的性能。针对以上问题，提出一种结合注意力机制的雷达多信号特征动作识别方法。首先，使用配置时分复用模式的多输入多输出毫米波雷达采集动作回波，将回波处理成短时能量、频率质心、相位变化(水平、俯仰)四维时序信号特征；然后，根据信号特征数据结构设计了多信号序列融合分类网络，该网络由1DCNN对信号抽取高维特征，再将特征送入GRU以充分提取时序规律，并引入Attention机制对重要特征映射加权赋予GRU隐含状态不同的权重，最终通过SoftMax层完成动作分类；最后，在实际采集的雷达多信号数据集上进行实验，结果表明，多信号序列特征可以充分表征人体动作，所设计的网络收敛速度快，对8种不同的动作分类，平均正确率达到了98.5%。更多还原

【Abstract】 The radar activity feature spectrum has poor characterization effect on wide human actions such as walking and falling. The single action feature and mismatch feature data structure can reduce the performance of activity recognition. Aiming at above problems, a radar multi-signal feature extraction method combining attention mechanism is proposed. Firstly, such method uses the multiple input multiple output millimeter-wave radar with time-division multiplexing mode to collect the action echo which is processed into four-dimensional time series signal features of short-term energy, frequency centroid and phase change(horizontal, pitch). Then, a multi-signal sequence fusion classification network is designed according to the signal feature data structure. In this network, 1 DCNN extracts high-dimensional features from signals, and then sends the features to GRU to fully extract timing rules. Besides, the Attention mechanism is introduced to map important features and assign different weights to hidden states of GRU,and action classification is completed through SoftMax layer. Finally, experiments on the radar multi-signal data set collected in practice show that the multi-signal features sequence can fully characterize human actions, and the designed network has a fast convergence speed. The average accuracy of 8 different actions is 98.5%.更多还原