SFSTrack: Spatial–Frequency Synergistic Transformer for Robust Nighttime UAV Object Tracking

Abstract

With the rapid advancement of unmanned aerial vehicle (UAV) technology, UAV-based visual object tracking has emerged as a significant research focus, particularly in the domain of remote sensing. However, low-light conditions at night severely degrade the accuracy and robustness of tracking algorithms, limiting the effective deployment of UAVs in nocturnal environments. Most existing trackers are primarily designed for normal lighting conditions and struggle to extract discriminative representations of degraded target features in low-light scenes. To address these challenges, we propose a novel spatial–frequency synergistic Transformer tracking (SFSTrack) model. The key of this model lies in the design of a spatial–frequency feature synergist (SFFS), which mainly consists of two components: a multilevel spatial residual enhancement (MSRE) block and a frequency-aware adaptive feature refinement (FAFR) block. On one side, the spatial layer improves features in dark regions under low-light conditions through iterative learning. On the other side, the frequency-domain refinement branch is constructed by integrating the Fourier transform, allowing the model to suppress noise interference while also enhancing high-frequency details. In addition, a confidence-weighted temporal feature fusion (CWTF) mechanism is designed, which adaptively integrates reliable target appearance information from historical frames, thereby reducing tracking drift caused by feature degradation in the current frame or large changes in target appearance. Experiments on public nighttime tracking benchmarks, including NAT2021, NAT2021L, DarkTrack2021, and UAVDark135, show that the proposed algorithm outperforms existing state-of-the-art methods, thus verifying the effectiveness and robustness of SFSTrack for UAV object tracking in nighttime low-light conditions.