Research On The Methods Of Radar Image Target Recognition And Tracking Based On Deep Learning

Radar plays a vital role in military reconnaissance and early warning,disaster rescue,environmental monitoring,topographic mapping,autonomous driving,and remote sensing.Through multi-dimensional data processing,radar images obtain the twodimensional or three-dimensional characteristics of the radar target to implement recognition and tracking.Typical radar images mainly consist of radar spatial images,Doppler spectrums,etc.With the development of electronic and information technology,radar can obtain continuous multi-frame image(such as video-SAR),to represent the variations of target characteristics over time clearly and accurately,and has been applied to recognition and tracking of slow-moving ground and ocean targets.Due to the significant differences between radar image characteristics and optical image characteristics,how to process the radar sequence quickly and reliably with high precision,and acquire the genre,motion state,intention and other information of a target is the key in radar image target recognition and tracking.Deep learning technology mainly exploits big data network training,which solves the problem of non-convex optimization for complex nonlinear models.It has been successfully applied to the fields of audio and video data analysis,and image processing.In recent years,with the rapid development of deep learning,it provides a possible new path for radar image target recognition and tracking.To this end,this dissertation is based on the deep neural network to research on the difficulties and hot spots in SAR image target recognition,micro-Doppler radar human behavior recognition,ground slow-moving target tracking.The main research contributions and innovations include:1)Targeting at the difficulty of manual annotation of radar images,a binary classification method under an unsupervised framework is proposed,i.e.feature growth training.Inspired by competitive learning,the proposed binary clustering approach for optical and radar image datasets is composed of five stages: feature elimination,feature seeding,feature germination,feature growing and feature grafting.Compared with the recent deep learning clustering methods including joint unsupervised learning of deep representations and image clusters(JULE)and deep embedded clustering(DEC),feature growth training is able to achive similar clustering performance in MNIST,fashion MNIST and MSTAR datasets.2)Aiming at the problem of human behavior recognition,deep learning based human behavior recognition methods are proposed.At first,the short-time Fourier transform(STFT),S transform(ST)and smooth pseudo-Wigner-Willey(SPWVD)characteristics of human behavior are analyzed.Secondly,based on the two-dimensional convolutional neural network(2DCNN)and long-short term memory structure(LSTM),a single spectrum based recognition method is proposed.On top of it,a multispectral behavior recognition method based on three-dimensional convolutional neural network(3DCNN)is proposed.Finally,combining the range information,a human behavior recognition method is proposed.Experimental results show that the recognition accuracy rate of the multi-domain joint method reaches 97.58%,which improves by 2.33% compared to the multispectral 3DCNN recognition method.3)Aiming at the difficult problem of video-SAR slow moving target detection and localization,a video-SAR moving target tracking framework based on slow-moving target SAR image shadow characteristics is established.It first obtains the registered video SAR image by video back-projection(v-BP)algorithm and realizes shadow tracking of video SAR by using a fully convolutional network(Siamese network).Then,combining the obtained tracking trajectory with the moving target back-projection(mBP)algorithm,accurate imaging of the ground moving target is acquired.Simulated experimental results show that the accurate reconstruction of the moving target can be achieved by m-BP when the velocity estimation error is less than 0.1 m/s.4)A fast multi-target tracking network for video-SAR is proposed.The network,based on the optical multi-target tracking framework Fair MOT,enables feature enhancement by adding attention mechanism to the detection backbone network.Ensuring the tracking performance,the tracking network capacity is significantly reduced,and the tracking efficiency is improved by optimizing the network structure.The measured data analysis shows that by combining the backbone networks DLA-34,DLA-18 and the fast multi-scale feature extraction module(FMs FEM)with the triplet attention mechanism,performance of shadow tracking for video SAR can be significantly improved.Through the analysis of tracking efficiency,we find that the frame rates of the proposed lightweight video SAR multi-target tracking network embedded with triplet attention mechanism are much faster than those of DLA-34 and DLA-18.For Sandia video SAR dataset with an image size of 1088×608,the average frame rate of FMs FEM and FMs FEM embedded with triplet attention mechanism reaches up to 60.32 fps and56.13 fps,respectively,which is about three times higher than the other networks.In conclusion,this dissertation takes radar image as the research object.By combining deep learning framework with radar signal characteristic,this dissertation studies the problem of radar image target recognition and tracking,and improves the performance of ultra wide-band radar image sequence target recognition and video-SAR target shadow tracking,along with the ability of target feature extraction and recognition,which provides meaningful explorations for future development of intelligent radar technology.