EHTracker: Toward Fine-Grained Localization for Satellite Video Target Tracking

Abstract

Deep learning (DL)-based methods have shown great potential in the satellite video target tracking community. Nevertheless, most of the methods are still plagued by poor localization capabilities and background interference problems. In this article, we propose EHTracker, a Siamese-based tracker based on enhanced head network, specifically for tiny targets in satellite videos. EHTracker enhances the localization capability for tiny objects through the proposed enhanced head network, while the proposed background suppression module (BSM) improves the network's ability to suppress background interference. Specifically, to alleviate the problem of overlapping between the foreground and background in the classification branch, we propose a multihead graph attention refinement module (MGARM). Second, to enable the regression branch to determine the location of feature-poor small targets more accurately, we introduce a gradual regression strategy (GRS). MGARM and GRS together constitute the enhanced head network component of EHTracker. Further, in order to work with the enhanced head network to achieve classification and localization of small targets that lack distinctive features, we design a BSM. Extensive experiments are implemented on the SatSOT dataset and SV248S dataset, and the experimental results show that our method achieves state-of-the-art tracking performance.

Keywords

• Graph attention
• satellite video tracking
• siamese network
• tiny object