Abstract
Infrared small-target tracking is a fundamental yet challenging task in computer vision, owing to extremely small target sizes, severe background clutter, and low signal-to-noise ratios, which often cause tracking drift or complete target loss. As a representative method in this field, SiamDT has achieved promising performance on the challenging Anti-UAV410 benchmark—a large-scale thermal infrared dataset. However, SiamDT still suffers from three key limitations: first, its standard Feature Pyramid Network (FPN) lacks channel-wise selective focusing, causing target-related channels to be overwhelmed by numerous background clutter channels; second, its fixed receptive field cannot adapt to the drastic scale variations in infrared targets; third, its spatial-domain feature fusion lacks high-frequency component compensation, resulting in severe loss of discriminative edge and contour details in deep features. To address these issues, we propose Enhanced-SiamDT, an enhanced tracking model built upon the SiamDT framework. Our contributions are twofold. First, we design a Multi-Attention Feature Pyramid Network (MA-FPN) that sequentially integrates Efficient Channel Attention (ECA), Large Selective Kernel (LSK), and Wavelet Domain Attention (WDAM) to achieve channel-wise recalibration, adaptive multi-scale receptive field selection, and frequency-domain detail enhancement, thereby suppressing background clutter and strengthening small-target feature representations. Second, building upon the dual similarity learning architecture inherited from the SiamDT baseline, we introduce a background prototype suppression strategy that reduces false alarms by penalizing candidate boxes with high similarity to background prototypes and a conservative template update mechanism with explicit update criteria, which prevents template drift under fast motion and short-term occlusion. Extensive experiments on the Anti-UAV410 and Anti-UAV benchmarks demonstrate that Enhanced-SiamDT achieves new state-of-the-art performance, with a State Accuracy (SA) of 68.58% and 71.84%, and a Precision of 89.92% and 92.71%, respectively. These results validate that our approach effectively overcomes the limitations of existing Siamese trackers, delivering significant improvements in discriminative feature extraction and tracking robustness for infrared small targets.