Research On High Precision Detection And Tracking Method Of SAR Moving Target

Synthetic aperture radar(SAR)moving target detection and tracking have important and wide application prospects in both military and civilian fields,such as illegal vehicle localization and tracking,road traffic monitoring,and intelligent vehicle navigation.However,SAR moving target detection and tracking still face some challenges.For SAR moving target detection,it is difficult to separate slow moving targets from background clutters,especially in the cases of high-speed platforms and strong clutters.The estimation errors of moving target velocity and position are large,and the target focusing effect is poor,which make slow moving target detection faces great challenges.Traditional SAR moving target detection methods use the unique advantage of moving target shadow to improve detection accuracy.However,shadow features extracted by traditional methods are not sufficient,resulting in false alarms and missed detections especially under the complex scene.Fortunately,deep learning can extract shadow features more adequately so as to improve moving target detection accuracy.For SAR moving target tracking,deep learning methods have obvious advantages over traditional methods in terms of tracking performance,but they still face some problems.When tracking targets of interest,existing deep learning methods ignore the target association information between frames,resulting in poor network transition.Targets of interest tracking performance will also be greatly reduced.When tracking multiple targets,existing deep learning methods suffer from low accuracy in estimating the target number.The tracking process is insensitive to the number change and the target identity switching.Therefore,these methods are prone to the missed and wrong tracking,which greatly reduce SAR multi-target tracking accuracy.Therefore,in order to address the above problems,this dissertation has carried out relevant research,and the main innovations and contributions are as follows.1.Aiming at the limited accuracy problem of slow moving target detection,SAR slow moving target detection based on four-antenna paired cancellation model is proposed.Firstly,the four-antenna paired cancellation model is established.It equips the front and rear antennas with the positive and negative oblique angle to increase the baseline length,then the displacement of the slow moving target is increased,which is beneficial to distinguishing the slow moving target from the background.At the same time,the model uses the attached antenna channels and the displaced phase center antenna(DPCA)algorithm to suppress background clutters in SAR images,so that the information of slow moving targets can be highlighted.Secondly,a refocusing iterative optimization(RIO)algorithm is proposed to improve the azimuth velocity estimation accuracy.It initially estimates the azimuth velocity using the displacement of the moving target in SAR images,and then the moving target phase error is compensated by iterative optimization,which improves the focus depth.The experimental results show that the proposed method can achieve high accuracy detection of SAR slow moving targets among strong clutters,and the accuracy of azimuth velocity estimation is improved by about two orders of magnitude compared with other methods.2.Aiming at the problem that the feature extraction of moving target shadows is inadequate,SAR moving target detection based on shadow detection network(Shadow De Net)is proposed.Firstly,the network uses a histogram equalization shadow enhancement(HESE)pre-processing technique to enhance the shadow saliency.Secondly,the transformer self-attention mechanism(TSAM)is used to weight different features to strengthen the shadow and suppress false alarms.Then,the shape deformation adaptive learning mechanism(SDAL)is used to focus on the shadow deformation to improve the robustness of the shape changes and reduce missed detections.Finally,the network uses the online hard example mining(OHEM)mechanism for selecting typical difficult negative samples to further suppress false alarms.The experimental results show that the detection accuracy of Shadow De Net is higher than that of other advanced networks by 9.00%.3.Aiming at the poor network transition problem of target-of-interest tracking,SAR moving target tracking based on the guided anchor siamese network(GASN)is proposed.Firstly,GASN uses many paired of images to train the siamese network,then the matching function to represent the similarity can be learnt,and the correlation information of the target between frames can be established.After that,for tracking target-of-interest,the semantic guided anchor-adaptive mechanism(SGAAL)uses the target features in the first frame as the priori information,and generates some anchors that better match the target shape in the next frame,which can effectively suppress false alarms.Finally,the similarity matching function is used to track the anchors most similar to the target-of-interest as the tracking result,thus improves the network transition.The experimental results show that GASN can achieve target-of-interest tracking,and it outperforms other advanced networks in terms of the network transition and tracking accuracy.4.Aiming at the problems of inaccurate target number estimation,wrong tracking and missed tracking,SAR multi-target tracking based on shadow-enhanced feature self-attention and anchor-adaptive network(SE-FSA-AA-Net)is proposed.Firstly,the network uses a sparse low-rank noise decomposition shadow enhancement(SLRNDSE)pre-processing technique to suppress noise interference and reduce missed detections.Then,the TSAM is used to weight different features to strengthen the shadow and suppress false alarms.Finally,the SGAAL is used to adaptively generate some anchors with position probability to filter out the background interference,and reduce estimation errors of targets number.The experimental results show that SE-FSA-AA-Net can estimate the number of targets more accurate than other advanced networks,and greatly reduce the probability of wrong or missed tracking,thus significantly improve the tracking accuracy by 6.40%.