ZenTrack: zero-cost template learning for robust thermal infrared small-target tracking

Abstract

Thermal infrared (TIR) small-target tracking is needed for continuous monitoring, such as in anti-UAV defence and wildlife observation. TIR imaging is a low-texture and feature-poor type that has a large homogeneous background, thus reducing the signal strength of the target. Dynamic-template frameworks are suitable for addressing time-varying changes, but they have serious defects in TIR scenarios: random template updates cause continuous accumulation of error drift, and excessive background content in the cropped patch reduces the contrast of the target. ZenTrack is a new transformer-based tracker proposed in this paper that can learn a strong template through training-only regularisation and is also zero-cost for online inference. Strictly inference-agnostic strategies are used only in the training stage to enhance the robustness of tracking without adding computational overhead or structural complexity at online deployment. Template Perturbation-based Robust Learning (TPRL) is added to vary templates dynamically during training to reduce the network's dependence on temporal updates and prevent drift caused by distractors. At the same time, the Target-Centric Template Learning (TCTL) module is also spatially constrained asymmetrically. Strictly apply this to the template branches and bypass the search region to suppress latent non-target interference, thus achieving a highly discriminative, target-focused representational space. Many experiments on the Anti-UAV410 and BIRDSAI benchmarks show that ZenTrack outperforms typical baseline methods. Compared with the baseline without TPRL and TCTL, ZenTrack has improved the AUC by +1.84% and precision on Anti-UAV410 by +2.93% without increasing online inference cost.