Abstract

Accurate and efficient road obstacle detection is crucial for ensuring the safety and reliability of autonomous and assisted driving systems. However, current object detection models, including YOLOv8, often encounter difficulties in accurately detecting small obstacles and balancing detection precision with inference speed. To address these challenges, we propose an improved YOLOv8-based detection framework featuring three key enhancements: (1) the backbone is redesigned using RepVGG modules to accelerate inference through structural re-parameterization without compromising feature representation; (2) an additional small-object detection head at the P2 feature level is introduced to significantly enhance sensitivity towards small-scale obstacles; (3) a Swin Transformer module is integrated into the final stage of the backbone to improve global contextual understanding and semantic feature extraction. Extensive experiments conducted on a comprehensive, multi-source road obstacle dataset demonstrate that our proposed model achieves superior performance with a precision of 0.812, recall of 0.782, mAP@0.5 of 0.822, and mAP@0.5:0.95 of 0.588. Additionally, it maintains a real-time inference speed of 180 FPS on an NVIDIA H800 GPU. The results confirm the efficacy of our approach, particularly in enhancing small object detection in complex real-world traffic scenarios.