Deep Learning-Based Object Tracking in Satellite Videos: A comprehensive survey with a new dataset

As a fundamental task for research in satellite videos (SVs), object tracking is used to track the target of interest in traffic evaluation, military security, and so forth. The current satellite technology in the remote sensing field makes it possible to track moving targets with a relatively high frame rate and image resolution. However, objects under this special view are often small and blurry, making it hard to extract deep features effectively. As a result, quite a few deep learning (DL) methods were proposed for object tracking in SVs. In addition, evaluation criteria for daily life videos (DLVs) are not fully applicable to SVs, which always get low precision evaluation results for tiny objects. In this article, we make three contributions to the research on SVs. First, a new single object tracking (SOT) dataset, named SV248S, is proposed, including 248 sequences with high-precision manual annotation, and 10 kinds of attribute tags are designed to completely represent the difficulties during tracking. Second, two high-precision evaluation methods are proposed, especially for small object tracking. Finally, 28 DL-based state-of-the-art (SOTA) tracking methods, from 2017 to 2021, covering popular frameworks, are evaluated and compared on the proposed dataset. Furthermore, some guidelines for effectively adopting DL-based methods are summarized based on comprehensive experimental results.