| Due to the advantages of simple structure,easy implementation and low cost,millimeter-wave radar is more and more widely used in civil areas,but it usually does not have the ability to recognize targets.Therefore,it is of great significance and has broad application prospects to realize automatic target recognition(ATR)based on radar by studying radar echo data.In order to make the millimeter-wave radar have the ability of ATR,the following researches are carried out based on 24 GHz linear frequency modulated continuous wave(LFMCW)radar in this paper:(1)For the target recognition of LFMCW radar,the feature extraction and selection based on radar echo data are studied.Firstly,the echo model of linear frequency modulated wave is constructed,the one-dimensional range profile(RP)of target is extracted by fast Fourier transform(FFT)from the fast time dimension of the echo.The fast and slow time two-dimensional FFT was used to separate multiple targets and estimate the velocity,range and other information of targets to extract the range-Doppler matrix(RDM)of targets.And the two-dimensional constant false alarm rate(2D-CFAR)algorithm is applied to the RDM data,which detects the target.Based on the relationship between radar cross-section(RCS)and the echo amplitude,maximum amplitude sequence in RDM is extracted as RCS feature.Aiming at the false alarm problem in the 2D-CFAR detection results,it is proposed to use the tracked target information to extract the amplitude and range-Doppler-time(RDT)data of the moving target from the original RDM.Research and analysis show that RCS sequence features are only suitable for a single scene because of the scale sensitivity.RDT feature solves the scale sensitivity problem of the RCS sequence and also contains the structure information of RP,so it is suitable for radar target recognition.(2)To construct a radar RDT feature data set.The radar data of vehicles in different weather environments such as daytime,nighttime and rainy day are collected.And the RDT feature of each target is extracted from the radar echo data,then the target category is labeled.Since there may be non-target data in the automatically extracted RDT data,these data need to be preprocessed.The preprocessing includes data cleaning and data standardization.The purpose of data cleaning is to remove dirty data such as other target data and empty data.Standardization is to unify data scale and remove dimensionality.Finally,the RDT feature data set of vehicle is constructed with the time dimension length of 15 frames and the RDM size of 15×15.(3)To build a target recognition model based on radar RDT data sets.According to the RDT feature of targets,a corresponding three-dimensional Convolutional Neural Network(3D-CNN)structure is designed.By comparing and analyzing the recognition effects of different structures,the best-performing 3D-CNN model is determined.Finally,based on the determined network structure,the recognition accuracy of radar RDT data with different overlapping degrees are compared.The result shows that when a group of features are extracted every 5 frames in RDT as input,the recognition effect is the best. |
