Training-Free Zero-Shot Aerial Tracking via Multi-Scale Foundation Correlation and Geometry-Locked Keypoint Correspondence

Training-free trackers built on frozen vision foundation models need no labelled video, but Vision Transformer features exist only on a coarse patch lattice. On unmanned aerial vehicle (UAV) footage, where targets are small and change scale continuously, this lattice limits box estimation rather than localization. We present Blihat, a fully training-free single-object tracker that decouples where the target is from how large it is. A frozen DINOv3 ViT-L/16 localizes the target by foreground-weighted correlation of a three-level feature memory over a three-scale search region. A SuperPoint+LightGlue stage then matches a scale-normalized crop against the frame-1 anchor and up to two keyframes whose correspondences are mapped back to frame-1 coordinates, so a single RANSAC similarity transform carries the exact initial box into the current frame without accumulated drift. The fitted box is used only when it passes a reliability gate that includes a scale lock bounding any size change to 1.44x, and it is fed back into the feature memory. On the full UAV123 benchmark, Blihat raises success AUC from 0.400 to 0.490 (+22.5% relative) and precision from 0.614 to 0.690 over an otherwise identical coarse-only tracker (paired t=6.1, 95% bootstrap CI [+0.062,+0.118]), at 4.6 FPS on an NVIDIA A100. On the held-out DTB70 benchmark, with no re-tuning, AUC improves from 0.383 to 0.420; on the long-term UAV20L benchmark it improves from 0.287 to 0.338, which is not statistically significant. A fine-only baseline with the same matcher, gate and memory but without DINOv3 reaches only 0.311 AUC on UAV123 even after tuning, so the gain comes from pairing the two stages. On UAV123, Blihat exceeds published hand-crafted correlation-filter trackers, but it remains below trained trackers, and below correlation-filter trackers on DTB70 and UAV20L, where non-rigid targets and long-term re-detection remain open problems. Code: https://github.com/DagimawiMoges/Zero-shot-UAV-object-tracking-with-coarse-to-fine-architecture