A Multi-Scale Context Fusion Network With Boundary Refinement for Road Marking Segmentation

Accurate segmentation of road markings, particularly symbol-type patterns, is vital for robust perception in Intelligent Transportation Systems (ITS) and autonomous driving. These markings provide essential visual cues for vehicle localization, navigation, and traffic regulation. However, the complex geometries and diverse visual appearances of symbol markings pose persistent challenges to current deep learning models, which often emphasize lane-line extraction while neglecting fine-grained boundary precision. To overcome these limitations, this paper proposes a multi-scale context fusion network with edge refinement designed for precise road marking segmentation. The network incorporates a Multi-Scale Semantic Enhancement Module (MSEM) to reinforce semantic coherence across scales and a PointRend inspired refinement head (PR-Head) to adaptively optimize boundary representations in uncertain regions. A focal loss function (FLF) is further introduced to alleviate the imbalance between foreground and background, leading to more stable and accurate predictions. Extensive experiments on the CeyMo dataset demonstrate that the proposed method achieves a mIoU of 72.6%, surpassing the baseline by 3.4% while maintaining real-time inference at 118.9 FPS. These results confirm that the network effectively enhances multi-scale feature representation and boundary precision with minimal computational cost, providing a practical and efficient solution for fine-grained road perception in complex traffic environments.