DAA-YOLOv8s with downsampling aware high resolution fusion for real time tiny vehicle detection in UAV imagery

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Detecting very small vehicles in UAV imagery remains challenging due to severe scale variations, dense object distributions, diverse viewing directions, and the loss of fine-grained spatial information caused by repeated downsampling. To address these limitations, we propose DAA-YOLOv8s (Downsampling-Aware Architecture for YOLOv8s), a lightweight detection framework that explicitly preserves high-resolution spatial information for improved tiny-object localization in aerial scenes. The proposed model integrates a Multi-Receptive Field Downsampling (MRFD) module, a C2f-Partial Residual Bottleneck (C2f-PRB), and a High-Resolution Rebalanced Fusion (HRRF) neck to preserve complementary spatial cues, reduce redundant computation, and strengthen high-resolution multi-scale feature representation. By operating the detection head on \(160\times 160\) , \(80\times 80\) , and \(40\times 40\) feature maps, the framework retains fine-grained structural and boundary information that supports reliable localization of vehicles under varying viewing directions, scales, and partial occlusion. Experimental results on the VisDrone2019-DET dataset demonstrate that DAA-YOLOv8s achieves 52.70% mAP@50 and 32.20% mAP@50–95, representing improvements of 11.10 and 7.60 percentage points, respectively, over baseline YOLOv8s. At the same time, the proposed model reduces the parameter count from 11.13M to 3.29M, corresponding to a reduction of approximately 70.4%, while maintaining comparable computational complexity, with GFLOPs decreasing slightly from 28.70 to 28.60. In real-time deployment, DAA-YOLOv8s achieves an end-to-end latency of 11.8841 ms and a throughput of 84.15 FPS, exceeding the 25 FPS input stream used in the deployment experiment. These results demonstrate that the downsampling-aware design provides a favorable balance between tiny-object localization capability, model compactness, and practical real-time inference. Furthermore, DAA-YOLOv8s achieves 80.89% mAP@50 and 57.50% mAP@50–95 on the VEDAI dataset, providing additional validation on an aerial vehicle detection benchmark. Overall, the proposed framework offers an efficient architecture for tiny vehicle detection in UAV imagery, with potential applicability to real-time aerial surveillance and intelligent transportation systems.