AUTHOR=Zhang Xiaoyu , Huang Hongwei , Peng Xiting , Zhang Xiaoling , Xu Lexi , Yang Yang 

TITLE=StaBle-MambaNet: structure-aware and blur-guided lane detection with Mamba

JOURNAL=Frontiers in Artificial Intelligence

VOLUME=Volume 8 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2025.1687983

DOI=10.3389/frai.2025.1687983

ISSN=2624-8212

ABSTRACT=The perception system constitutes a critical component of autonomous driving, due to factors such as high-speed motion and complex illumination, camera-captured images often exhibit local blurring, leading to the degradation of lane structure clarity and even temporary disappearance of lane markings, which severely compromises the accuracy and robustness of lane detection. Traditional approaches typically adopt a two-stage strategy of “image enhancement followed by structural recognition” Initially, the entire image undergoes deblurring or super-resolution reconstruction, followed by lane detection. However, such methods rely on the quality of full-image restoration, exhibit low processing efficiency, and struggle to determine whether the disappearance of lane markings is genuinely caused by image blurring. To address these challenges, this paper proposes an Inter-frame Stability-Aware Blur-enhanced Mamba Network (StaBle-MambaNet), which identifies blurred regions and assesses the presence of potential lane structures without relying on full-image restoration. The method first localizes blurred areas and employs a Structure-Aware Restoration Module to perform directional extrapolation and completion for potential lane line regions. Subsequently, the Blur-Guided Consistency Reasoning Module evaluates structural stability to identify genuine lane regions. Finally, enhanced features are constructed into a spatially continuous token sequence, which is fed into a lightweight state-space model, Mamba, to model the dynamic feature variations in blurred regions while preserving the vertical structural evolution of the image. Experimental results demonstrate that StaBle-MambaNet significantly outperforms existing mainstream methods across multiple public lane datasets (e.g., CULane and CurveLanes), particularly under challenging conditions such as nighttime, occlusion, and curved lanes, exhibiting clear advantages in both detection accuracy and structural stability.