Research On Human Classical Motion Classification In Low Frequency Radar Based On Deep Learning

The radar sensor has the advantages of all-time and all-weather working,non-contact and so on,so it can be used for non-invasive reconnaissance and surveillance of the human body.And this kind of sensor has gradually become an important method of sensing in urban anti-terrorism and remote health monitoring.The low-frequency radar can capture the motion information of the human body through the wall,but its microDoppler time-frequency information is blurred,which is hard to distinguish.The rise of deep learning also makes the low-frequency radar has the potential of recognizing the visually indistinguishable micro-Doppler time-frequency features of some human motions.Based on the low-frequency MIMO SFCW-UWB radar,this paper focuses on the motion classification of human target walking,sitting,and falling behaviors based on the CNN model.This paper first studies the dataset acquisition of the low frequency radar.According to the Boulic human body model,it establishes the radar echo model with SFCW as the transmitting signal,and then establishes the basic processing flow of the radar signal.The micro-Doppler time-frequency analysis method is compared,and the time-frequency feature extraction of human target in different data domain is studied.The dataset acquisition based on low frequency radar is realized by using simulation and measured data.Secondly,the paper studies the methods of the radar dataset preprocessing and the dataset expansion.Based on the preprocessing method of dataset in deep learning,a preprocessing method suitable for radar time-frequency image dataset is selected.Two kinds of dataset expansion methods are proposed based on two-dimensional timefrequency domain scale transform and sub-band micro-Doppler difference.The effectiveness of the data set expansion method is verified by simulation,measured data and training results on CNN.Finally,the paper studies the classification of human motion in low frequency radar based on CNN model.According to the framework of the CNN model and the principle of each layer,the human motion classification task of low-frequency radar is realized on the Alex Net.Two CNN optimization design methods based on model whole-layer compression and single-layer structure compression are proposed.Based on the model optimization,the RMRnet model was obtained,which achieved a classification accuracy of 99.35% on the measured dataset.