Vehicle detection in remote sensing images faces significant challenges due to small object sizes, scale variation, and cluttered backgrounds. To address these issues, we propose RSW-YOLO, an enhanced detection model built upon the YOLOv8n framework, designed to improve feature extraction and robustness against environmental noise. A Restormer module is incorporated into the backbone to model long-range dependencies via self-attention, enabling better handling of multi-scale features and complex scenes. A dedicated detection head is introduced for small objects, focusing on critical channels while suppressing irrelevant information. Additionally, the original CIoU loss is replaced with WIoU, which dynamically reweights predicted boxes based on their quality, enhancing localization accuracy and stability. Experimental results on the DJCAR dataset show mAP@0.5 and mAP@0.5:0.95 improvements of 5.4% and 6.2%, respectively, and corresponding gains of 4.3% and 2.6% on the VisDrone dataset. These results demonstrate that RSW-YOLO offers a robust and accurate solution for UAV-based vehicle detection, particularly in urban scenes with dense or small targets.

This paper proposes an optimized algorithm based on YOLOv11s to address the problem of insufficient detection accuracy of vehicle targets from a drone perspective due to certain scenes involving complex backgrounds, dense vehicle targets, and/or large variations in vehicle target scales due to oblique imaging. The proposed algorithm enhances the model’s local feature extraction capability through a module collaboration optimization strategy, integrates coordinate convolution to strengthen spatial perception, and introduces a small object detection head to address target size variations caused by altitude changes. Additionally, we construct a dedicated dataset for urban vehicle detection that is characterized by high-resolution images, a large sample size, and low training resource requirements. Experimental results show that the proposed algorithm achieves gains of 1.9% in precision, 6.0% in recall, 4.2% in mAP@0.5, and 3.3% in mAP@0.5:0.95 compared to the baseline network. The improved model also achieves the highest F1-score, indicating an optimal balance between precision and recall.