Person identification with limited training data using radar micro‐Doppler signatures

Radar‐based human micro‐Doppler analysis has been the subject of much investigation in recent years. Apart from the conventional activity classification task, person identification based on human movement signal has emerged as a research interest. This paper presents a method to recognize a person's identity from varied human motions using an ultra‐wideband radar. The human movement data is captured in an indoor environment and is then transformed into micro‐Doppler spectrograms for identification. Moreover, as it is always challenging to construct large scale radar datasets in practice, we adopt a plain convolutional neural network with a multi‐scale feature aggregation strategy to address the identification problem. Experimental results show that the micro‐Doppler signatures have great potential in person identification, and our model presents relative satisfying performances limited training set. Especially, when “walking” is used for identification, our approach achieves a person identification accuracy of 96.8% for the four targets used.