A Joint Moving Target Detection Method in Video SAR Via Low-Rank Sparse Decomposition and Transformer
Hui Fang,Guisheng Liao,Yongjun Liu,Cao Zeng,Xiongpeng He,Mingming Xu
DOI: https://doi.org/10.1109/jstars.2024.3503639
IF: 4.715
2024-01-01
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Abstract:Video synthetic aperture radar (SAR) has exhibited considerable potential for detecting and tracking ground moving targets. Numerous classical shadow-based detection methods have been applied in video SAR. Additionally, shadow-assisted detection methods based on convolutional neural networks (CNNs) have been developed. In this paper, we propose a joint detection method for moving targets in video SAR, which can combine the information of the video SAR image and the corresponding sparse image to suppress background interference sufficiently and improve detection accuracy. Specifically, low-rank sparse decomposition (LSD) technology is first applied for video SAR images to generate their corresponding sparse images in which the background is eliminated and shadows of moving targets are enhanced. Then, we improve Faster RCNN and build a two-stream extraction feature network based on the Transformer structure that allows the video SAR image and the sparse image as input simultaneously as well as extracts and fuses the features from two types of the images, which can acquire more discriminative target features, thereby improving the final the detection performance. Furthermore, the improved Faster RCNN only modifies the original feature extraction network. Thus, it can adopt the same training and test manner as Faster RCNN, greatly facilitating its utilization. Finally, experiment results on Sandia National Laboratories (SNL) data demonstrate the proposed detection method outperforms other state-of-the-art methods. And our method reduces the false alarms by 1.02%, the missed detections by 43.24%, and increases the mean average precision (mAP) by 2.98%.