Abstract:

In the task of satellite video object tracking,the performance of the existing approaches is restricted due to the low resolution of the target,background clutter and occlusion.In this work,a new satellite video object tracking method based on multi-granularity learning and motion state estimation is proposed.The multi-granularity learning applies the bi-directional fusion network to adaptively fused shallow features and the deep features,which enhances the representative ability of the fused features with the rich spatial information from the shallow features and the strong semantic information from the deep features.Moreover,the motion state estimation utilizes the historical movement state of the target to estimate the locations of the target in the current frame,and refines the movement state outputted by the tracking network,which improves the robustness of the tracker.Finally,a new satellite video object tracking algorithm based on the two proposed methods is presented,and evaluated on the satellite video object tracking dataset,SatSOT.Experimental results reveal that the proposed tracker achieves a better performance than the other trackers.The proposed tracker surpasses the Siamese-based tracker and SiamCAR by 5.1% and 3.2% on the precision score and the success score,respectively.

Key words: satellite video, object tracking, multi-granularity information learning, motion state estimation, siamese network