DCD-YOLO: An Improved YOLOv11n Algorithm \for Traffic Participant Detection

Detection of traffic participants is one of the core tasks in environmental perception for autonomous driving. However, accurately identifying small sized and occluded objects in complex traffic scenarios remains a pressing challenge. To address these issues, a high precision multi-scale perception architecture named DCD-YOLO is proposed based on YOLOv11n, which integrates multi-branch feature extraction, attention guided enhancement, and task decoupled prediction mechanisms.A Dynamic Fusion Block (DFB) is introduced to partially replace the original C3k2 structure, enhancing feature extraction for small objects and complex backgrounds. In addition, the Content Guided Attention (CGA) module is incorporated in the Neck to emphasize key object regions through the collaborative effects of channel, spatial, and pixel level attentions.Meanwhile, a Dynamic Shared Enhanced Head (DSEH) combines reparameterizable detail enhanced convolutions, shared convolutional weights, and Group Normalization to improve multi-scale object localization and classification performance while maintaining a lightweight architecture.Experimental results demonstrate that DCD-YOLO improves mAP@0.5 by 2.9% on the KITTI dataset. Furthermore, the proposed model is deployed on a real vehicle platform and tested under realistic driving scenarios, which further verifies its deployment feasibility and practical significance for autonomous driving systems.