Wavelet-enhanced transformer for object tracking in satellite videos
Meiyu Chen, Peng Wang, Xue Wu · Journal of Physics Conference Series · 2025
Abstract Aiming at the problems of limited tracking performance and insufficient adaptability faced by current Transformer-based visual target tracking methods when dealing with scenes of small target size and blurred features in satellite videos, this study proposes a Wavelet-Enhanced Transformer for Object Tracking in Satellite Videos (WAETrack). This model addresses the issue of the Transformer’s insufficient ability to model local spatial features. By leveraging the multi-scale characteristics of the wavelet transform, a local enhanced attention module for the wavelet transform is constructed. Through enhancing the high-frequency subband features decomposed by wavelet transform, the model’s perception ability for the local spatial features of the target is improved. Provide richer discriminative features for tracking targets in complex scenes. To explore the effectiveness of the algorithm, experiments are conducted on the OOTB dataset, composed of 110 video sequences in this study. The results show that WAETrack has better performance compared with other tracking algorithms, achieving an accuracy rate of 70.65% and a success rate of 52.57%. It is also verified that using wavelet transform to capture local features and high-frequency details can effectively improve the tracking performance of the model.