YOLOv11‐MSRC: A Multi‐Scale Reconfigurable Framework for Real‐Time Road Safety Detection

ABSTRACT

Road traffic accidents cause approximately 1.19 million deaths worldwide annually, highlighting the need for reliable visual detection systems in intelligent transportation. Existing object detection methods often struggle under challenging conditions such as haze, low light, dense traffic, and occlusion, while highly complex models restrict practical deployment in latency‐sensitive applications. To address these challenges, this study proposes YOLOv11‐MSRC, a multi‐scale reconfigurable framework for road safety visual detection. The framework integrates three components: Re‐parameterised Feature Pyramid Network (RepFPN), a multi‐scale re‐parameterised convolution module that strengthens feature fusion during training and merges into a single dense branch before deployment; Lite Shift Head (LSH), a shift‐based decoupled detection head to reduce task interference between classification and regression; and improved Intersection over Union (iXIoU), a loss with dynamic weighting to emphasise hard samples and improve localization robustness. Experiments on visible‐light and infrared traffic datasets show consistent gains over the YOLOv11n baseline, with mean Average Precision (mAP)@0.5 gains of 4.85 and 2.60 points and mAP@0.75 gains of 5.62 and 7.92 points, respectively. YOLOv11‐MSRC retains real‐time GPU inference with moderate computational overhead.