Unmanned Aerial Vehicles (UAVs) are essential for traffic monitoring in Intelligent Transportation Systems (ITS). However, the detection of traffic objects in drone-view imagery still faces serious challenges due to drastic scale variations, dense small objects, and complex background clutter. Existing detectors often struggle to effectively separate targets from background noise and lack efficient cross-scale feature interaction mechanisms. This paper introduces ED-DETR, a refined real-time Detection Transformer specifically designed to optimize performance in UAV scenarios.
Specifically, a Dual-Stream Coordinate Attention (DCA) module is designed, which can effectively extract significant texture features while preserving context information through parallel maximum pooling and average pooling operations. In addition, this paper introduces an efficient multi-scale attention structure in the feature fusion neck, which combines grouping convolution and cross-spatial learning to significantly improve the positioning ability of small targets. A large number of experiments on the VisDrone2019 dataset show that ED-DETR achieves an mAP of 40.4%, which is significantly higher than that achieved by RT-DETR and other state-of-the-art detectors. By harmonizing high detection accuracy with low computational overhead, the proposed framework enables effective real-time monitoring of aerial traffic scenarios.