<scp>TCTGNet</scp> : A Real‐Time Object Detection Method for Dense Traffic Scenes in Aerial Photography

ABSTRACT Addressing challenges in aerial dense traffic scenes, such as small object scales, dense distributions, severe occlusion, and strong background clutter, existing efficient detectors for aerial scenes often exhibit limited multi‐scale representation and unstable localization, hindering an effective trade‐off between accuracy and real‐time performance. To this end, we propose TCTGNet, an efficient real‐time object detector tailored for aerial dense traffic scenarios. Built on a single‐stage real‐time detection framework, TCTGNet improves dense small‐object detection through coordinated adaptation from three perspectives: enhanced feature representation, key information selection, and training optimization. Experiments on VisDrone2019 show that TCTGNet achieves 51.6% mAP over 10 traffic‐related categories, while running at 128 FPS with 11.3M parameters, yielding a better balance between accuracy and complexity than mainstream methods. Moreover, results on UAVDT further confirm its adaptability and stability under different data distributions. Overall, TCTGNet offers a practical solution that jointly achieves high accuracy and real‐time performance for urban traffic perception, while showing promising deployment potential on UAV‐related resource‐constrained platforms.