GEN-THERMAL-UAV: A SYNTHETIC BENCHMARK FOR MULTISCALE FIXED-WING UAV TRACKING IN THERMAL INFRARED VIDEO

The detection and tracking of fixed-wing Unmanned Aerial Vehicles (UAVs) in thermal infrared (TIR) imagery constitutes a critical, yet data-scarce domain in modern computer vision and airspace security. Fixed-wing UAVs operate at higher speeds and altitudes than rotary-wing counterparts, presenting a unique “approaching” threat profile that transitions rapidly from a sub-pixel thermal anomaly to a resolved object. Existing datasets fail to capture this full dynamic range. This paper proposes Gen-Thermal-UAV, a novel synthetic dataset generated using the Gemini Veo 3 video diffusion model. We utilize a seed-driven approach: the dataset of 220 videos (1,760 seconds) is derived from only two real thermal images. This pipeline distills the thermodynamic fidelity of real sensors into diverse, auto-labeled synthetic videos using SAM 2 for zero-shot annotation. We present a structured prompt engineering framework for scientific video synthesis and verify this dataset as a benchmark specifically dedicated to the multiscale approach-following task in thermal air-to-air engagements.