Research On Person Identification Based On Radar Micro-doppler Spectrogram

Person identification plays a vital role in preventing terrorist attacks and identifying illegal intrusions and has broad application prospects in military fields,home automation,human-computer interaction,and medical monitoring.Researchers have shown an increasing interest in radar-based person identification due to its advantages of robustness to the environment and protection of visual privacy.In this thesis,ultra-wideband radar is used to collect radar echo signals of moving human targets to generate micro-Doppler spectrogram,and then the characteristics of microDoppler spectrogram of human targets are analyzed,the micro-Doppler characteristics of radar to carry out the research on person identification.The main research work includes:This thesis proposes a person identification model based on multiscale feature fusion because of the multi-scale person identification features in the radar micro-Doppler spectrogram.The characteristics of different receptive fields of different deep convolutional neural networks are used to realize pedestrian recognition feature extraction of different scales in micro-Doppler spectrograms.The multi-scale features extracted by the convolutional neural network are fused to obtain the general features representing the person identification.Finally,the person identification is completed through the fully connected layer.By analyzing the characteristics of the micro-Doppler spectrogram of human motion,and aiming at the problem with large intra-class variance and small inter-class variance in person identification under various actions,this thesis proposes a person identification based on radar micro-Doppler contrastive learning model.Using the contrastive learning method to realize radar micro-Doppler feature comparison,extract fine-grained features from the micro-Doppler spectrogram,and improve the accuracy of person identification.Further improvement of the person identification performance is achieved by combining the micro-Doppler multi-scale feature fusion method.In terms of model optimization,by combining the Cross-Entropy loss function and Kullback-Leibler divergence,a more compact class distribution is obtained by promoting the separation of similar samples between classes and reducing the distance between samples with large differences within classes.The effectiveness of the proposed person identification method is proved by comparing a variety of existing person identification algorithms and image recognition algorithms on the radar-measured data set.The generalization performance of the person identification model proposed in this thesis is analyzed on the data sets with different noise levels,shorttime Fourier transforms window lengths,and the number of recognized people.Experiments show that the method in this thesis has good recognition performance and generalization performance in the person identification task based on the radar micro-Doppler spectrum.