Robust Online Satellite Video Object Tracking with Self-adoption Uncertainty

Ziye Wang,Duoqian Miao

DOI: https://doi.org/10.1007/978-3-031-65668-2_6

2024-01-01

Abstract:Tracking moving targets in satellite videos has lately gained popularity. However, the growth of target tracking in satellite videos is slower than in general videos due to the following factors. Satellite video tracking faces challenges stemming from low frame rates, causing significant object movement between frames and impacting prediction accuracy, while high-resolution footage exacerbates tracking difficulties by requiring extensive search regions for targets occupying a minimal percentage of the total pixel count. In overall, the level of uncertainty surrounding the target in the satellite footage is excessive. To address the above problems, we propose a novel DiMP-based tracker and introduce SES to stabilize the estimation of target motion in satellite videos. Furthermore, we introduce an uncertainty measure, which is included into network outputs and a loss function to remove unreliable samples from the training set. Extensive experiments have demonstrated that our technique can track targets with the highest precision score and success rate on both the SatSOT and SV248S datasets, achieving state-of-the-art status.