Research On SAR Target And Radar Point Clouds Target Recognition Based On Deep Learning

Target recognition plays an important role in the field of radar,among which Synthetic Aperture Radar(SAR)imaging and target recognition is essential to military and civil industries.Traditional SAR can only get two-dimensional images.Nowadays,radar imaging gradually expands from two-dimensional to three-dimensional.Among them,point cloud radar can provide rich information and lay a foundation for highprecision target recognition.Millimeter-wave MIMO(Multiple Input Multiple Output)radar has shown great potential in the field of 3D point cloud due to its high resolution in distance,angle and low cost.In order to fully utlize the potential of radar target recognition algorithm based on artificial intelligence,this thesis develop MIMO millimeter wave 3D point cloud radar,and applies the point cloud recognition network to the data of this radar.The main research work of this thesis is:(1)SAR target echo modeling and imaging are presented.The thesis establish airborne SAR echo model and derive echo spectrum.The classical imaging algorithm is introduced.Then the thesis analysis aircraft target scattering center,and obtain simulated aircraft SAR image.(2)The SAR images classification using machine learning and convolutional neural networks are presented.Classification experiments are performed using the simulated aircraft dataset and the public vehicle dataset MSTAR.In order to make the network adequately trained,the thesis use DCGAN to generate samples.Experiments show that network pre-train by generated images tend to have better performance.(3)The MIMO millimeter-wave radar 3D point cloud imaging and point cloud data processing are introduced.This thesis uses the millimeter-wave radar developed on the AWR1843 platform to collect three-dimensional point clouds of pedestrians,non-motor vehicles,and automobiles.Annotation for semantic segmentation are performed on the dataset.In order to integrate the data collected by multiple frames or multiple devices,point cloud registration is necessary.This thesis introduces the RANSAC and ICP algorithms,which have high registration accuracy for millimeter wave 3D point clouds.(4)Three point cloud recognition algorithms,Point Net,Point Net++ and Point CNN are presented.The point cloud recognition algorithm is applied to the data of MIMO millimeter-wave radar.The classification and segmentation experiments are also carried out on Model Net40 and Semantic KITTI public datasets to assist in the evaluation of algorithm performance.