In object detection for traffic scenarios, YOLO-v11 suffers from critical deficiencies in small-object detection: insufficient extraction of fine-grained information leads to low recall rates for targets like bicycles and tricycles, poor capture of global semantics causes frequent false and missed detections in complex backgrounds, and its redundant backbone network makes deployment on resource-constrained devices difficult. To address these issues, this study proposes three targeted optimizations: first, adding a 160 × 160 high-resolution prediction layer to construct a four-scale detection system for enhancing small-object detail-semantic fusion; second, embedding a Multi-scale Self-attention Module (MSAM) in the Neck layer to focus on small-object regions and suppress background interference. Validation on the VisDrone2019 dataset shows that the improved model significantly boosts performance: small-object recall and precision rates increase by an average of 24.3% and 13.7%, respectively; overall mAP50 and mAP50-95 rise by 7.5 and 9.6%, while parameters only increase by 8.2%. Ultra-small object detection rate reaches 71.4%, effectively supporting intelligent transportation tasks such as non-motor vehicle statistics and accident early warning.