MSF-ULDNet: A multi-scale feature fusion network for urban lane line detection

Abstract

This study addresses the issues of unstable lane line detection at both near and far distances in urban road scenarios and poor performance under interference factors such as lighting conditions, obstructions, and blurred road markings in the LSTR algorithm. This study proposes the Multi-Scale Feature Fusion Network for Urban Lane Line Detection (MSF-ULDNet). Based on the LSTR model, at the feature extraction level, it integrates a lighter-weight, efficient multi-scale attention mechanism (DS-EMA) and an adaptive feature pyramid network (AFPN) to obtain semantically rich, high-quality multi-scale feature representations. DS-EMA optimizes the spatial distribution of feature maps to mitigate the conflicting effects of complex interference information on lane detection, while AFPN enhances the model’s ability to represent lane lines at different distance scales by combining the detailed information of low-level features with the semantic information of high-level features. At the representation level, a new lane line shape prediction method is proposed, introducing a U-shaped curve as compensation in the original lane line model, providing the model with richer geometric constraints and enhancing its adaptability to lane lines with different curvatures. Experimental results show that, while maintaining low computational cost, the proposed method achieves a 0.54 percentage point improvement in accuracy metrics on the Tusimple dataset, a 2.91 percentage point increase in F1 score on the CULane dataset, and a 2.27 percentage point gain in IoU on the BDD100K dataset.