Vehicle detection in drone aerial views based on lightweight OSD-YOLOv10

Abstract To address the challenges of low performance in vehicle image detection from UAV aerial imagery, difficulties in small target feature extraction, and the large parameter size of existing models, we propose the OSD-YOLOv10 algorithm, an enhanced version based on YOLOv10n. The proposed algorithm incorporates several key innovations: First, we employ online convolutional reparameterization to construct the OCRConv module and design a lightweight feature extraction structure, SPCC, to replace the conventional C2f module, thereby reducing computational load and parameter count. Second, we integrate an efficient dual-layer feed-forward hybrid attention module to enhance the model’s feature extraction capabilities. We also construct a dual small-target detection layer that combines shallow and ultra-shallow features to improve small-target detection. Finally, we introduce the DySample dynamic upsampling module to enhance feature fusion in the neck network from a point sampling perspective. Extensive experiments on the VisDrone-DET2019 and UAVDT datasets demonstrate that OSD-YOLOv10 achieves a 40.7% reduction in parameter count and a 3.6% decrease in floating-point operations, while improving accuracy and mean average precision by 1.3% and 1.6%, respectively. Compared to other YOLO series and lightweight models, OSD-YOLOv10 exhibits superior detection accuracy and lower computational complexity, achieving an optimal balance between high accuracy and low resource consumption. These advancements make it particularly suitable for deployment in UAV onboard hardware for vehicle target detection tasks. Code will be available online (https://github.com/Z76y/OSD-YOLO).