Research On Human Behavior Recognition Based On Fmcw Radar

In the research of human behavior recognition using radar sensor,the traditional method depends on the micro Doppler effect of human motion and machine learning.By constructing the micro Doppler time-frequency analysis diagram of motion process as the input,the characteristics of each action are extracted manually,and then sent to the machine learning model to judge the output.However,the micro Doppler effect of human motion is easily affected by strong trunk echo,environmental noise,and changeable posture.The extracted feature map is weak,fuzzy and unstable,and the distinction effect of different behaviors is not intuitive and obvious.Moreover,the manual extraction of features seriously depends on experience,which is inefficient and has limited generalization ability.In order to solve the above problems,this thesis proposes a radar human behavior recognition method based on continuous frame accumulation range Doppler image processing and deep learning model.The main work is as follows:(1)This thesis introduces the frequency modulated continuous wave radar and its hardware system,expounds its ranging and velocity measurement principle,and makes a theoretical analysis of the machine learning model and depth learning model used in this thesis.(2)From the perspectives of time domain and distance domain,the traditional time-frequency analysis and the range Doppler processing based on frame accumulation are carried out to obtain the time-frequency analysis diagram and the range Doppler heat map of frame accumulation.Nine kinds of human behavior data sets are constructed respectively.The feature extraction method based on PCA dimensionality reduction is adopted,and then human behavior recognition is carried out through machine learning model,which verifies the superiority of the processing method based on frame accumulation range Doppler analysis proposed in this thesis.(3)Human behavior recognition based on convolutional neural network is proposed.The Efficient Net model is used to directly recognize the range Doppler heat map based on frame accumulation proposed in this thesis,and the most ideal recognition model in this thesis is found in Efficient Net-B4 model;Compared with the traditional convolutional neural network VGG16,Res Net50 and the model input by time-frequency analysis method,the experiment shows that the recognition effect of frame accumulation range Doppler analysis processing method and Efficient Net-B4 model is better.(4)Human behavior recognition based on transfer learning is proposed.Based on the Efficient Net-B4 model that completes parameter pre training in Image Net data set,the network model structure of human behavior recognition is constructed,and the secondary training is carried out through the radar data set made in this thesis.The experimental results show that compared with the model without transfer learning,the accuracy of nine kinds of human behavior recognition has been greatly improved,and the generalization ability is stronger.