High-precision human activity classification via radar micro-doppler signatures based on deep neural network

Radar-based human activity recognition has been of great interest due to its capability to resolve problems of the security and health system. Deep learning-based methods are widely used to recognize human motion at a micro-scale. However, most of the deep learning networks require large amounts of data. Here, we propose high-precision and efficient human activity classification method via radar micro-doppler signatures with data augmentation and deep neural networks. The proposed method can achieve higher than 99% classification accuracy for different human micro-motions. The most useful solution for classification accuracy improvement is the data augmentation and we try different ways and finally two effective methods are chosen, i.e., selecting different rangebins and spectrogram amplitude display values. In the network model, we compared the recognition accuracy of our model, AlexNet and VGG16 in human activity classification, and found that VGG16 has better generalization ability. In data augmentation, we compared the impact of different rangebins and different display amplitudes on recognition accuracy during human activity classification. Experimental results show that the accuracy deviations generated by selecting different rangebins and spectrogram display amplitude values for target classification are about 2.56% and 1.31% respectively. Selecting the optimal parameters to expand the data can achieve higher than 99% classification accuracy. It is demonstrated that selecting the appropriate rangebins and setting the optimal spectrum display amplitude are crucial for processing micro-doppler signals of the raw radar data.