Object Classification Methods Of 3D Radar Images Based On Deep Learning

With millimeter wave(MMV)near-field imaging systems applied at many secure locations such as airport security,researches on target classification and detection for three-dimension(3D)radar images have been of great theoretical value and commercial value.3D radar images contain abundant 3D spatial information and radar imaging features.This paper attempts to improve the ability of target classification for 3D radar objects by means of deep learning,making full use of the characteristics of 3D radar images.Chapter 1 introduces the research background and significance of this project,summarizing the state-of-the-art development of the field related to our dissertation,“Object classification methods of 3D radar images based on deep learning”,in detail in three sections.Subsequently,problems to be solved have been declared,followed by the main structure of this article.Chapter 2 explains the work of data preparation for the training and testing of networks,which includes the acquisition of simulation data and actual imaging data.It tells the principle of graphical electromagnetic computing and main steps to obtain simulation results,resemble to 3D radar images.Then the acquisition of imaging data is is explained,and then comparisons between simulation data and imaging data are carried out to verify the effectiveness and usability of the simulation data.Materials described in this chapter provide data basis for training and testing of methods discussed in following chapters.Chapter 3 studies the deep learning classification method for 3D radar targets based on two-dimension(2D)images.The idea,multi-view,is introduced to combine2 D features from each viewpoint of 3D objects.Through the feature-level data fusion,feature stretched from single-view image forming a global descriptor through max-output process.When it is applied to the projections of the iso-surface image of 3D radar image in multiple directions,good classification results are obtained on simulation data when the number of viewpoint reaches 80,and it also maintains stable classification performance on actual imaging data.Chapter 4 studies the deep learning classification method for 3D radar targets based on the input of point cloud.Points with 3D coordinates go through a hierarchical feature learning network,which includes sampling,clustering,feature extraction by point in one abstraction operation.Then,all rested points are combined to obtain a global descriptor for this point cloud to be classified.Compared with the image-based method,this network based on point cloud shows an advantage in capturing local and subtle structure,and it achieves better classification performance of complex targets than method based on 2D images.Besides,this chapter also discusses the influence of the non-uniformity of 3d radar data on classification results and preliminarily analyzes the feasibility to classify radar objects by network trained on data generated by 3d model without simulation data.In addition,in order to make full use of the rich information of 3D radar images,this chapter proposes a classification method that encodes intensity information into point cloud,and verifies its classification performance.Chapter 5 gives a conclusion of the whole dissertation,pointing out the main contributions,analyzing and comparing various methods proposed in this paper.Then insufficient places of this research are listed,also with some ideas for the future work.