Abstract: With the rapid development of intelligent transportation systems, real-time and accurate detection of abnormal objects on roads has become a key challenge in the field of autonomous driving safety. Addressing the issues of shape variability, scale diversity, complex environment, and overlapping targets in the task of road abnormal object detection, based on the YOLOv8n framework, three improvements are proposed at the model architecture level. First, the DCNv3 (Deformable Convolution Network v3) is introduced to replace the C2f module, thereby enhancing feature extraction capability. Second, the GAM (Global Attention Mechanism) is incorporated to establish cross-channel interactions and strengthen the representation of critical features. Finally, the AFPN (Asymptotic Feature Pyramid Network) is adopted to optimize multi-scale fusion and improve detection accuracy for multi-scale objects. Comparative and ablation experiments conducted on the COCO fusion dataset demonstrate that the improved model outperforms both the baseline YOLOv8n and other mainstream models in terms of precision and mAP (mean Average Precision). Compared with YOLOv8n, the proposed model achieves improvements of 5.7% in precision and 8.6% in mAP. Visualization results further indicate that the proposed improvement strategies effectively reduce both false positive and false negative rates, enhance the model's discriminative ability for road anomaly features, and improve its robustness in complex scenarios.