<p>Thermal infrared (TIR) target tracking benefits significantly from multi-scale feature fusion, especially in challenging scenarios such as low illumination, adverse weather, and complex environments. However, existing methods usually rely on feature pyramid architectures and often suffer from semantic inconsistency across scales in infrared imagery, which leads to noisy responses and inaccurate localization. In this paper, we propose a progressive feature ensemble tracker for TIR target tracking. Feature responses are first grouped into different scale levels according to their spatial resolutions. Multiple responses within each scale are then fused using Kullback–Leibler divergence to suppress noise and improve response quality. To exploit complementary information across scales, we further introduce a coarse-to-fine translation estimation strategy, in which higher-scale responses provide semantic guidance and lower-scale responses preserve spatial details. In addition, a response-confidence-guided adaptive filter update mechanism is developed to improve robustness against appearance variations. Experiments on the PTB-TIR, LSOTB-TIR, and Anti-UAV410 benchmarks show that the proposed tracker achieves competitive performance under challenging conditions such as occlusion, scale variation, and background interference.</p>