Improved RT-DETR Algorithm for Small Object Detection in Drone Aerial Imagery

Detecting small objects in aerial imagery presents unique challenges, including complex backgrounds, significant scale variations, and densely clustered targets. Conventional detection algorithms struggle with these conditions. An improved RT-DETR detection algorithm is introduced to address small object detection challenges in drone aerial scenes. The RT-DETR network architecture has been redesigned by incorporating three key innovations: a Global Boundary Propagator (GBP) backbone structure that extracts and preserves edge information in early network stages to enhance perception of small object boundaries; a Multipath Kernel Cascade (MKC) module that employs varied-scale parallel convolutions to improve feature representation; and a Cross Scale Feature Nexus (CSFN) network that incorporates bidirectional feature pyramid structures to enhance multi-scale feature fusion. Experimental results in the VisDrone dataset demonstrate that our proposed method achieves superior performance in small object detection with a AP<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50</inf> of 38. 1%, AP<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</inf> of 12. 6%, while maintaining real-time inference speeds of 54.14 FPS, providing an efficient and reliable solution for small object detection in drone aerial imagery.