SiamDA: a detail-attentive Siamese network with infrared optical saliency for pixel-level UAV tracking

With the rapid development of unmanned aerial vehicle (UAV) technology, establishing effective management systems for unmanned aerial vehicles has become increasingly important. Tracking small UAVs in complex environments using infrared imagery is a crucial yet challenging task, owing to limited target visibility and significant background clutter. Further, existing feature extraction methods struggle to effectively capture pixel-level infrared UAV signatures. Therefore, this paper introduces SiamDA, a detail-attentive anchor-free Siamese tracker designed to capture more infrared spectral details to enhance the representation of weak UAV targets. First, a detail-attentive network that employs deformable convolutions to capture fine-grained features, along with a Taylor-difference-inspired edge enhancement module to sharpen boundaries and reinforce geometric shapes of small UAVs. Then, a normalized Wasserstein distance loss and a dynamic template update scheme are integrated to improve tracking robustness. Evaluations on public near-infrared UAV datasets indicate that SiamDA attains an average precision (P₅) of more than 80%, surpassing state-of-the-art trackers trained on the same dataset.