Using Deep Learning To Detect Stationary Targets Of Automotive Millimeter-wave Radar

The automotive millimeter-wave radar is a mainstream sensor currently used to detect the distance,velocity and angle of the target ahead.Since the Doppler frequencies of the echo signals caused by different speeds are unequal,the automotive millimeter-wave radar can distinguish the echo and the clutter of the moving target by their different speeds.However,when the millimeter-wave radar and the target ahead are at rest,the echo of the front target does not generate Doppler frequency,and the collected echo data is a strong superposition of ground clutter and target signal so that the radar can't distinguish the target echo from the echo data easily,which is the main problem of the automotive millimeter-wave radar.In order to solve the problem of stationary target recognition of millimeter-wave radar,this thesis proposes a deep learning-based schema to identify stationary targets and detect the distance between the target in front and the radar.The research in this thesis takes 77G automotive millimeter-wave radar as object and by transmitting the radar signal to the front stationary target and then receiving the target reflected echo,the beat signal is processed after mixer processing,and then performs the two-dimensional FFT(Fast Fourier Transformation)operation on the beat signal,consequently obtaining signal strength information of different speeds at different distances of all targets in front of the radar.Finally,the research extracts all the echo intensity information with velocity 0,and uses the RNN and one-dimensional CNN to identify the existence of the stationary target as well as extracts the target distance.The main research contents are listed as follows:This thesis researches the problem that the automotive millimeter-wave radar is unable to identify the target when it is relatively stationary with the object.This thesis illustrates the application of FMCW in automotive millimeter-wave radar as well as how to use the beat signal of FMCW to extract the distance and speed of the targets. This thesis illustrates the basic structure and principles of machine learning,especially the CNN and RNN used in this thesis,and explains the neural network structure related with this thesis in detail.This thesis proposes a deep learning-based method to identify the stationary target of automotive millimeter-wave radar,which uses the RNN and one-dimensional CNN,and introduces the scheme and experimental design for the identification of stationary targets in detail,including data collection,preprocessing,feature extraction,and model design,training and verification.In the end,the thesis summarizes and analyzes the completed research work,and prospects the future research direction and future work.