Research On Shadow Enhancement And Tracking Method Of Video SAR Moving Target

Video SAR moving target detection and tracking technology is an important research direction in the field of SAR in recent years.However,due to moving target imaging position offset,change of illumination angle,etc.,its scattering characteristics are vulnerable to background interference and instability,which makes the moving target detection and tracking accuracy of video SAR low and even missed detection occurs.Shadow-based SAR moving target detection and tracking methods have attracted much attention because the target shadow reflects the true location,without offsetting and defocusing.However,it is not easy to distinguish the target shadow from the low scattering area in SAR images,and the target shadow features are not fully extracted and represented,which limits the accuracy of detection and tracking.In order to improve the accuracy of video SAR moving target detection and tracking,research on moving target shadow enhancement and tracking methods based on deep learning is conducted in this thesis.The main contents and contributions of this thesis are as follows:1.To solve the difficulty of distinguishing the shadow of moving target from low scattering background and extracting feature in video SAR,an enhancement method of shadow space structure decomposition based on low rank and sparsity and a robust enhancement method are proposed.Firstly,an enhancement method of shadow space structure decomposition based on low rank and sparsity（SEFV）is proposed.Its processes are: applying the video frame vector quantization on the video SAR image sequence;analyzing the signal space structure;establishing the low rank and sparsity decomposition model of video SAR signal with the characteristics of high inter frame correlation of video image background and low inter frame correlation of moving target shadow and small proportion of shadow pixels in the frame;constructing their regular terms for optimization processing to separate the background and target shadow,so as to reduce the background interference and enhance the shadow characteristics.The experiment results show that this method effectively enhances the shadow contrast and improves the detection accuracy.However,as the noise of video SAR image increases,the performance of SEFV algorithm decreases.Therefore,a shadow space structure decomposition enhancement method based on low rank sparsity and Gaussian distribution of noise（RSEFV）is proposed.This method analyzes the reason for the weak anti-noise performance of SEFV,which can’t suppress the noise interference in the shadow of moving targets only through sparse regularization.The noise obeying Gaussian distribution is introduced into the signal space model,and the influence of noise on shadow features is reduced by constructing regularization constraints on noise.The experiment results show that,compared with SEFV,RSEFV is performs better at anti-noise,shadow enhancement and detection accuracy.Especially when the signal-to-noise ratio is low,RSEFV is still better at shadow enhancement,showing the good robustness of the method.Finally,through the verification of multiple target detection networks based on deep learning,its detection performance has been improved,showcasing the universality of RSEFV.2.To solve the problem that low detection score of moving target shadow in video SAR results in tracking missing and affects tracking accuracy,a multi-target tracking algorithm based on adaptive spatial feature fusion and recall mechanism is proposed.Firstly,the problem that the target with low detection score in video SAR image has a great impact on the tracking accuracy is analyzed.The tracking performance of the multi object tracking algorithm（Byte Track）with recall mechanism is analyzed.The algorithm recalls the low detection score targets and matches them with the trajectory,which reduces the missed alarms.But the algorithm has obvious differences in the feature extraction of targets（shadows）with different scales,which affects the accuracy of shadow detection.To this end,the adaptive spatial feature fusion（ASFF）is combined with the multi-scale feature pyramid.In addition,a large number of low-score detected targets cause the increase of measurement noise in Kalman prediction model,leading to severe error in trajectory prediction.Therefore,the detection confidence is used to realize the adaptive adjustment of measurement noise and make the trajectory prediction more accurate.In addition,Gaussian smoothing interpolation is used to reduce the discontinuity of the trajectory and further optimize the tracking trajectory.The experiment results show that,compared with Byte Track,the ASFF and recall mechanism adopted by this method significantly reduces the false alarms and missed alarms and improves the tracking accuracy.In addition,the shadow robust enhancement method is combined with the methods in this chapter to further improve the tracking accuracy.