Research On Human Activity Recognition Method Of FMCW Radar Based On CNN

Human activity recognition has a wide range of needs in security monitoring,disaster relief,medical systems and other fields.Human activity recognition based on radar system has become one of the research hotspots in recent years because it is not affected by sight distance and light.At present,most of the research employ deep learning algorithms to classify according to range and micro-Doppler spectra in the field of human activity recognition of radar.However,the small datasets published in the current study are easily lead to over-fitting,which limit the application of deep learning in human activity spectra recognition.In order to extract useful information as much as possible from small radar datasets and train a model with superior recognition ability to realize effective human activity recognition,the following research work has been done in this paper:Firstly,the data augmentation method based on Generative Adversarial Networks(GAN)is explored to hand the problem of insufficient samples in small datasets.In this paper,GAN that has learned the range-time map(RTM)and Doppler-time map(DTM)data distribution is used to generate images,and the augmented training set is obtained by adding generated images.Extensive experiments have proved that the proposed data augmentation method effectively improve the generalization ability of the classifier in small radar datasets,and the recognition accuracy based on DTM is improved from 92.53% to 95.41% by training network with augmented training set.Secondly,this paper proposes a recognition method based on the fusion of range features and micro-Doppler features to solve the poor performance by using only one of them.This method inputs double-stream feature into the network instead of single feature,which enable the network to learn the characteristics of human activity in space distance and the micro-Doppler characteristics of the human body at the same time,so as to improve the feature extraction ability of the network.Compare with the single feature recognition method,accuracy of the human activity recognition method based on fusion features is improved by more than 1.5%.Thirdly,an end-to-end Cross Stage Partial-Fusion Convolutional Neural Network(CSP-FCNN)is designed for accurately identify human activity recognition by combining and optimizing the residual connection and the cross stage partial connection structure.CSP-FCNN uses CSP-residual connection structure to replace the ordinary convolution layer,which make it can fully extract features,so as to improve the accuracy of detection.The visual analysis of the feature maps shows that the CSP-FCNN with CSP-residual connection structure has outstanding learning ability.Further,abundant experiments are used to optimize the hyper-parameter of CSP-FCNN,and the best accuracy of CSP-FCNN is 97.67%.