SDU-RTDETR: a small object detection algorithm for UAV perspective based on RT-DETR

Embedded devices and computer vision are spawning many applications, object detection from UAV (Unmanned Aerial Vehicle) perspectives has become increasingly challenging due to issues such as insufficient multi-scale information, limited target extraction capabilities, and imbalanced scale distributions. To overcome these problems, we designed a new detection model, SDU-RTDETR, which is derived from the RT-DETR framework and has been optimized. Specifically designed for detecting small objects in UAV imagery, SDU-RTDETR introduces two key improvements: A novel backbone module, CSPHDRAB, inspired by the Cross Stage Partial (CSP) network, which significantly enhances the extraction of small object features; An improved neck network incorporating the Small Object Enhancement Pyramid (SOEP), which boosts the fusion of multi-scale information. According to our experimental findings on the VisDrone2019 dataset, SDU-RTDETR outperforms the baseline RT-DETR-R18 model by 4.8%, achieving a mAP50 of 50.6%. Because of these improvements, SDU-RTDETR is now a more potent option for tiny object identification in real time in UAV applications.