A General Target Tracking and Sample Generation Framework for Satellite Videos Leveraging on Geometry-Aware Low-Rank Representation

Video satellites observe the Earth with high spatial-temporal resolution, empowering remote sensing with better capability for emergency needs. The tracking of moving targets naturally has received a great deal of attention. Many methods are proposed and achieve satisfying results after years of research, where the low-rank representation (LRR) shows great potential due to its weak assumptions on data. However, the original LRR lacks the consideration of frame-wise geometric inconsistency led by imperfect video stabilization, significant view change, and elevation difference, resulting in pseudo-motion and declined performance on some video sequences. Therefore, we propose a geometry-aware LRR (Geo-LRR) that simultaneously considers the alignment and the target tracking in videos. Moreover, to relieve the dilemma of high-quality samples, we generate pixel-wise annotations with straightforward post-processing on sparse images. Extensive experiments demonstrate the effectiveness of our method with 2.051 pixel-wise RMSE compared with groundtruth.