HEATTrack: History Evidence Injection and Target-Aware Prompting for Aerial Multispectral Single-Object Tracking

Abstract

Aerial multispectral single-object tracking (SOT) is fundamental to inspection, search-and-rescue, and long-range surveillance from moving platforms. However, it remains highly vulnerable to temporal drift caused by tiny targets, fast platform motion, abrupt viewpoint changes, and cluttered backgrounds. Multispectral imagery (MSI) provides complementary reflectance cues beyond RGB, but reliably exploiting spectral information under band noise and background variability is still challenging. We present HEATTrack, a Transformer-based framework that balances short-term temporal evidence and long-term target identity cues for stable aerial multispectral tracking. First, history evidence injection (HEI) injects lightweight short-term temporal evidence by rasterizing recent predictions into boxmaps and fusing them with selected-band historical observations, yielding evidence maps that suppress irrelevant background and provide observation-conditioned guidance under noisy historical predictions. The history evidence module is trained with a perturbation-and-annealing strategy to reduce the mismatch between clean training histories and imperfect inference-time predictions. Second, target-aware multitemplate prompting (TAMTP) preserves anchor-guided target identity cues by injecting compact structural and spectral signatures into multitemplate prompts and aggregating templates with reliability-aware weights, thereby strengthening identity discrimination under distractors. Extensive experiments on MUST and MSITrack demonstrate consistent gains over competitive multispectral and general-purpose baselines. On MUST, HEATTrack improves the area under curve (AUC) from 62.8 to 67.1 over our reproduced UNTrack* baseline. On MSITrack, HEATTrack achieves 56.2 AUC, improving over our reproduced UNTrack* by 5.7 points. The source code, trained models, and reproduction files will be released upon publication.