Reliability-Aware Lane Detection for Autonomous Driving in Complex Nighttime Environments

Abstract

To address the challenges of unstable lane line segmentation, incomplete structural representation, and insufficient reliability in complex nighttime traffic environments due to low illumination, glare, road surface reflection, and local occlusion, a reliability modeling method for lane line detection in autonomous driving scenarios is proposed. This method is based on semantic segmentation, integrating lane region identification and prediction reliability representation into a unified framework. It enhances effective semantic representation in nighttime road scenes through feature encoding and context aggregation. Based on this, a lane segmentation branch and a reliability learning branch are constructed to generate pixel-level lane responses and corresponding confidence information, respectively. To improve the continuous representation of slender lane structures in complex backgrounds, structural consistency constraints are further introduced to mitigate the adverse effects of boundary ambiguity, local missing values, and spurious response interference. Simultaneously, a confidence modulation mechanism is designed to guide the optimization of the segmentation results, ensuring that the final output maintains lane region discrimination capability while more effectively suppressing uncertain responses in nighttime scenes. Experimental results compare and analyze typical semantic segmentation models, demonstrating that the proposed method achieves superior overall performance in complex nighttime traffic environments, exhibiting outstanding performance in lane region identification accuracy, structural integrity, and result stability. Related research provides a technical approach for nighttime autonomous driving lane line perception that balances semantic representation and reliable modeling.