Abstract:As unmanned aerial vehicle (UAV) technology develops rapidly, UAV-based object detection has broad application prospects in fields such as smart city traffic monitoring. However, UAV-captured images face challenges such as large object scale variations and complex backgrounds, making it difficult for traditional detection algorithms to balance precision and efficiency. This study proposes an optimized UAV object detection method based on RT-DETR. To address the problem of large target scale variations in UAV-captured images, this study optimizes the multi-scale feature fusion strategy, and the algorithm’s ability to detect objects at different scales is enhanced by improving the P3–P5 feature pyramid structure. Additionally, considering that objects in UAV scenarios are typically small and densely distributed, the AIFI attention mechanism is optimized to improve the algorithm’s feature representation capability for small objects. Based on the lightweight RT-DETR-R18 design, the proposed method significantly reduces computational complexity while maintaining detection precision. The experimental results on the VisDrone dataset which contains UAV-captured images across ten categories demonstrate that compared to the original RT-DETR algorithm, the proposed method improves detection precision for small objects while maintaining real-time performance, validating the effectiveness of each improved module.