SDF-YOLO: A Multi-Scale Dynamic Feature Fusion Network for Enhanced Lightweight Object Detection in Autonomous Driving

To enhance the multi-scale object perception and representation capabilities of the YOLOv8n model in complex autonomous driving scenarios, this paper proposes SDF-YOLO, a lightweight object detection algorithm based on a multi-scale dynamic feature fusion mechanism. Building upon the original YOLOv8n architecture, several structural optimization strategies are introduced to strengthen its multi-scale feature modeling ability. First, an additional detection head is incorporated to process low-level, small-scale features, thereby improving the model's sensitivity to fine-grained objects. Second, four Sample, Parameters Free Attention-Adjacent Context Coordination Modules (Sim-ACCoM) are embedded in both the backbone and neck networks. These modules leverage convolutions with varying dilation rates to enhance contextual awareness and cross-scale information integration during feature extraction. Furthermore, DynamicInterpolationFusion (DIF) modules are introduced after each C2f module and before feature concatenation in the neck to enable efficient coupling and complementation of multi-branch features, thus improving the richness and discriminative power of the overall feature representation. Finally, a Group-Slim pruning algorithm is applied to remove redundant parameters, improving parameter efficiency and enhancing network stability. Experimental results on a self-annotated Car dataset containing 11 object categories demonstrate that SDF-YOLO outperforms the original YOLOv8n model, achieving improvements of 1.7%, 3.4%, 3.3%, and 3.7% in Precision, Recall, mAP50, and mAP50-95, respectively, while reducing model parameters by 1.2M. These results confirm that the proposed SDF-YOLO significantly improves detection performance in complex scale scenarios, enhances model robustness and accuracy, and exhibits strong practical potential for autonomous driving applications.