Moving Target Shadow Detection Based on Deep Learning in Video SAR
Hao Zhang, Zhe Liu
Abstract
Video SAR is a well-explored subject because of its high imaging frame rate and continuous monitoring of targets. Due to moving target shadow (MTS) is not defocused, it can reflect actual position of moving targets well, which makes it possible to detect the position of moving target by shadow. In our work, Feature pyramid network (FPN) is adopted as backbone in our method to extract multi-shape and multi-intensity features about MTS. Then the fused feature maps are fed to the traditional region proposal network (RPN) for generating proposals and regions of interest (ROIs). Next, separate classification and regression branches are employed. A novel box offset regression, dense local regression (DLR), is accepted for further improving the detecting accuracy of MTS, while the traditional classification is used to achieve final detection. On the dataset formed of the real data published by Sandia National Lab (SNL), the detecting accuracy of MTS by our method reaches 92.44% and a comparison with a traditional method based on OTSU is given.