Abstract
To address the challenges in vehicle detection from unmanned aerial vehicle (UAV) overhead images, such as small object size, low resolution, complex background, and scale variation, this paper proposes several targeted improvements to the YOLOv11n model. Firstly, inspired by the Cross Stage Partial Networks (CSPNet), a Dynamic Multi-Scale Edge Enhancement Network (DMS-EdgeNet) is designed to improve robustness to local target features. This module applies multi-scale pooling to extract edge features at various scales and dynamically fuses them through adaptive weighting. Secondly, the DynaScale Aggregation Network (DySAN) module is introduced into the neck network, and a multi-level jump connections structure is adopted to fuse low-level and high-level boundary semantics, thereby improving the detection capability of fuzzy boundary targets and improving target positioning accuracy under complex imaging conditions. Finally, a P2 small target layer is added to further improve the accuracy of small target detection. Based on these innovations, we propose a new architecture named Dynamic Multi-scale and Channel-scaled YOLO (DMS-YOLO), significantly improve the model's ability to perceive small targets. Experimental results show that DMS-YOLO improves mAP50 and mAP50-95 by 7.0% and 2.9%, respectively, on the Aerial Traffic Images dataset, and by 5.1% and 3.1% on the VisDrone-DET2019 dataset, demonstrating superior performance over the YOLOv11n baseline.