Radar Human Action Recognition Based On Improved Convolutional Neural Network

Human action recognition technology has important application significance in the fields of security and monitoring systems,monitoring of the elderly at home and hospital patients.The radar-based human action recognition technology is not affected by light,weather and other environments,is non-invasive,and the micro-Doppler effect generated by human movements reflects the kinematic characteristics,human action recognition can be realized by using its micro-Doppler characteristics.Convolutional Neural Network(CNN)can learn features by themselves,no need to manually set features,so they are widely used in the field of radar human action recognition.But the convolution kernels of traditional convolutional neural networks are simplex,and with the increase of the number of network layers,it is prone to vanishing gradients.Therefore,this thesis will conduct research on radar human action recognition technology based on improved convolutional neural network.The main content is as follows:1.Simulate and construct micro-Doppler spectrogram dataset of human actions.The design and simulation are carried out for the four actions of walking,boxing,arm swinging and hand waving,and then the radar echo signal of human movement is calculated,later,the time-frequency analysis of the radar echo is performed by Gabor transform to obtain the micro-Doppler spectrogram and construct the dataset for the training and testing of network model.2.A radar human action recognition method based on multi-scale features fusion convolutional neural network is proposed.This method introduces convolution kernels of multiple sizes into the conventional network model to extract and fuse deep features to realize the recognition of human actions.Since the extracted multi-scale fusion features contain local structural information of different sizes,it has better classification performance,thereby improving the recognition rate.The simulation results show that when the signal-to-noise ratio is 0d B,the average recognition rate of the method for four kinds of human actions is improved by 1.95% compared with the corresponding single scale feature network.3.A radar human action recognition method based on improved residual convolutional neural network is proposed.This method improves the conventional Residual Networks,removes the batch normalization layer,retains the original convolutional structure and residual structure,and reduces the scale and number of convolutional layers.On the one hand,it greatly reduces the time required for identification.At the same time,high recognition performance is also guaranteed.The simulation results show that when the signal-to-noise ratio is 0d B,compared with the18-layer Residual Networks,the average recognition rate of this method for 4 kinds of human actions is similar,and the network model parameters are only 6.18% of its parameters.The time required to identify a single human action spectrogram is only13.62% of the time required.4.A radar human action recognition method based on the convolutional neural network with integrating attention module and multi-scale features is proposed.This method integrates attention module in the multi-scale feature fusion convolutional neural network,focusing on the most effective channel convolution features and spatial local area convolution features for classification,thereby improving the recognition rate if human actions.The simulation results show that when the signal-to-noise ratio is 0d B,the average recognition rate of this method for four kinds of human actions is improved by 1.91% compared with that when the attention module is not integrated.