FTTrack: RGB-T Tracking With Frequency-Adaptive Fusion and Temporal Enhancement

Abstract

RGB-T tracking exploits complementary information from RGB and thermal infrared (TIR) modalities to achieve robust object tracking. Mainstream approaches exhibit limitations in two critical aspects: inadequate cross-modal feature fusion and insufficient temporal modeling, resulting in performance degradation during challenging scenarios such as rapid motion, occlusion, and significant appearance variations. To overcome these limitations, we introduce FTTrack, a novel RGB-T tracking framework incorporating two key components: Frequency-Adaptive Feature Fusion (FAFF) and Temporal Track Query Memory (TTQM). The FAFF module adaptively enhances high-frequency details in RGB features while preserving low-frequency contour information in TIR features through frequency-domain filtering, thereby facilitating more effective cross-modal interaction and complementary feature integration. The TTQM module employs learnable temporal query tokens to propagate and accumulate historical target states across consecutive frames, significantly enhancing temporal consistency and robust target representation. Comprehensive experiments conducted on the three benchmark datasets demonstrate that FTTrack achieves promising improvements over mainstream trackers.