DMC-YOLO: an improved algorithm for pedestrian and vehicle detection in foggy weather

To address the challenges of suboptimal multi-scale object detection accuracy and model parameter redundancy in foggy scenarios, this paper proposes an improved YOLOv11-based algorithm for pedestrian and vehicle detection in hazy conditions. By introducing the C3k2_DyConv module, the approach significantly enhances detection precision while streamlining model parameters. A Multi-Scale Feature Aggregation (MSFA) module is employed to hierarchically partition features into multiple levels, enabling each layer to independently capture rich information across different scales and semantics, thereby acquiring diverse contextual features and strengthening the model's perception and comprehension of complex foggy scenes. Furthermore, the integration of the CARAFE upsampling mechanism not only generates more accurate upsampled features but also dynamically adjusts the upsampling process according to input feature variations, effectively improving the model's generalization capability and facilitating more efficient processing of complex and diverse image data. Experimental results demonstrate that compared to YOLOv11s, the proposed algorithm achieves a 3.0% improvement in mean Average Precision (mAP) on the RTTS dataset, and shows a 4.7% enhancement over YOLOv6s on the RSUD20K dataset, exhibiting substantial potential for practical applications.