Reliability‐Guided RGB‐IR Fusion for UAV Traffic Detection

ABSTRACT

Small traffic targets in unmanned aerial vehicle (UAV) imagery are easily missed because of overhead viewpoints, dense distributions, occlusion and complex backgrounds. RGB images provide texture and colour cues but degrade under low‐light or shadow conditions, whereas infrared images are more illumination‐robust but often have blurred boundaries. This letter proposes ERDF‐Anchor, a lightweight reliability‐guided decision fusion method for UAV RGB‐infrared traffic detection. A four‐channel early‐fusion detector supplies primary anchors, while RGB‐only and infrared‐only detectors serve only as auxiliary experts for consistency verification, confidence reweighting, unsupported‐anchor filtering and weighted box refinement. Object‐level expert statistics on DroneVehicle show that 50.6% of objects are more reliable in RGB and 46.8% in infrared, confirming local complementarity. Under a unified post‐processing protocol, ERDF‐Anchor improves recall from 0.9141 to 0.9321 and mAP50 from 0.8197 to 0.8252 over RGBIR4‐Early, with less than 1 ms post‐processing overhead per validation image on saved predictions.