Research On Human Activity Classification Based On Micro-doppler Radar And Convolutional Neural Network

Micro-Doppler radar-based human activity classification can be used in many fields,such as sleep monitoring,elderly care,human-computer interaction,and counter-terrorism monitoring.However,there are still many deficiencies in the existing classification algorithms from radar data representation,feature extraction and classification recognition analysis.So far,almost all classification algorithms recognize radar-based actions by first performing short-time Fourier transform on the original data,and the original radar data is represented as a spectrogram.Then utilize manual methods or neural networks to extract features from the spectrogram,and finally use a classifier to classify.The two-step process is not an end-to-end neural network in the true sense.The short-time Fourier transform is a fixed transformation that represents the original radar data from the amplitude-time domain to the frequency-time domain.This process is a process of spectrum analysis of signals,not a transformation process specifically designed to classify radar data,so the optimal representation of radar data will be limited by this process.There may be representations that represent radar data into other spaces that are more conducive to the classification of actions,rather than timefrequency domain space.In this paper,we design an end-to-end neural network that combines three parts: data representation,feature extraction,and classification recognition.The first part of the network replaces the short-time Fourier transform process with two one-dimensional convolutional layers,making the feature representation process trainable,thereby training feature representation methods that are more beneficial to action classification,and improving the accuracy of classification.There are still some problems in identifying human actions based on micro-Doppler radar.The neural networks used for classification are two-dimensional convolutional neural networks.Although these methods have high classification accuracy,they have high computational complexity and high power consumption.These methods regard the spectrogram as a common optical image.As for radar data,each column of pixels in the spectrogram can be regarded as a one-dimensional time vector sequence,it may have much temporal correlations among columns.The pixels of an optical image have high spatial correlations.Therefore,it is not the best choice to treat the two-dimensional spectrogram as an ordinary optical image and then use a twodimensional convolutional neural network for classification.One-dimensional neural networks have a better ability to extract time-related information.In this paper,we design a fully onedimensional convolutional neural network for micro-Doppler radar data.The network contains multi-scale technology,dense connection technology and channel compression technology to improve the performance of the network.It not only improves the recognition accuracy of actions based on radar data,but also keeps the computational complexity at a low level.