AEMFOD: Asymptotic Elimination Multi-Scale Fusion for Real-Time UAV Traffic Detection

Object detection is a fundamental perception capability for UAV-based intelligent transportation systems, underpinning applications including traffic monitoring, incident detection, and collision warning. However, reliable detection from aerial imagery remains difficult because UAV scenes typically contain multi-scale objects, diverse viewpoints, imbalanced object distributions, and stringent onboard computational constraints. Moreover, most existing studies primarily emphasize offline accuracy improvements, often overlooking the requirements of real-time, resource-constrained deployment. To simultaneously consider detection performance and onboard efficiency, we develop an Asymptotic Elimination Multi-scale Fusion Object Detection framework, termed AEMFOD, for UAV traffic imagery. Specifically, an asymptotic elimination feature pyramid network (AEFPN) is introduced to reduce computational complexity, while the asymptotic weighted coordinate fusion (AWCF) module operates within AEFPN to strengthen the integration of multi-scale features and preserve informative low-level representations. The proposed detector is evaluated through both benchmark experiments and practical UAV flight experiments. Compared with the YOLOv5s baseline, AEMFOD achieves relative $A{P}_{50}$ improvements of 4.8% on DOTA and 31.5% on VisDrone. Furthermore, embedded experiments quantify inference speed, and flight experiments yield qualitative detection samples under practical UAV operating conditions, showing competitive latency and a compact model footprint suitable for onboard traffic applications.