TDF-YOLOv8 : An Enhanced UAV Object Detection Method Based on Triplet Attention and Dynamic Scaling

In this paper, challenges in object detection based on UAV imagery are addressed. The focus is on high-resolution imagery, varied perspectives, and dense arrangements of small objects. The YOLOv8s model is enhanced by integrating Triplet Attention into its C2f module to improve recognition precision of densely distributed targets. A novel Detect_Dyhead structure is introduced to dynamically adjust detection strategies across different scales and shapes. Additionally, the FocalerShapeIoU loss function is employed to refine bounding box accuracy. Experimental results on the VisDrone2019 dataset show that Precision, Recall, [email protected], and [email protected]:0.95 metrics are enhanced by 2.1%, 1.4%, 1.4%, and 1.2%, respectively, while the model size is reduced by 0.5MB. Strong performance is also demonstrated on the WiderPerson and custom Tanks datasets, underscoring the generalizability of the proposed TDF-YOLOv8 model.