Research On Shadow Detection And Tracking Technology Of Video SAR Moving Target

Video SAR technology is one of the research hotspots in the current SAR field.Compared with traditional SAR,video SAR can obtain continuous image sequences of the monitoring area at high frame rate and high resolution,achieve high frame rate microwave imaging effects,and intuitively reflect the changes in the monitoring area scene and target motion state in a dynamic way.The current video SAR system mainly performs unified imaging on the monitoring scene,and does not perform motion compensation and focusing on the moving target.The dark shadow formed by the movement of the target can reflect the real position and state information of the target.Therefore,it is of great significance to study the shadow characteristics of video SAR targets and how to effectively use shadows to detect and track moving targets.This thesis conducts research on moving target detection and tracking technology based on video SAR shadow characteristics.The main research contents and contributions are as follows:1.The imaging principle and shadow characteristics of video SAR in circular bunching mode are studied.The main factors affecting the imaging frame rate are deduced and analyzed through mathematical formulas,and a mathematical model is established to analyze the formation principle and characteristics of shadows in video SAR image sequences.The similarities and differences of shadows formed by stationary targets and moving targets are compared,indicating that shadows in video SAR can reflect the real position of targets,and the feasibility of using shadows to detect and track moving targets is analyzed.2.A video SAR data set that can be used for detection and tracking is constructed.To solve the problem that currently there are few public video SAR data,especially the video SAR data set based on deep learning,this thesis uses the video SAR data made public by SANdia National Laboratory and another measured video SAR data to construct a video SAR deep learning data set that can be used for detection and tracking.3.The shadow detection method of video SAR Based on deep learning is studied.Aiming at the problems of complex detection process and manual design of components in fast RCNN and Yolo detection networks,which are the most commonly used and representative detection networks in SAR field,this thesis proposes a video SAR shadow detection framework based on transformer codec network.The framework has no anchor design and NMS post-processing,which simplifies the detection process.The performance of shadow detection is improved by using image histogram equalization,multi-layer feature fusion,deformation convolution and improved channel attention module.The proposed method achieves good shadow detection results on SNL video SAR data and additional video SAR data.4.The video SAR multi-target tracking method with joint detection and data association is studied.For existing multi-target tracking methods,only high-scoring detection results are usually screened and simple data association with Kalman filter prediction results leads to poor tracking performance.This thesis uses a low-score detection frame association algorithm and builds a video SAR multi-target tracking framework.The experimental results on SNL video SAR data verify the effectiveness of the proposed multi-target tracking method.In addition,experiments on video SAR simulation data show that improving video SAR image quality can effectively improve the performance of moving target shadow detection and tracking.