<p>Lane detection, a critical technology for autonomous driving systems, directly impacts driving safety through its reliability. However, challenges such as lane diversity, occlusions, and illumination variations in complex road scenes impose stringent demands on detection robustness. In this paper, we propose an enhanced lane detection model, which can reinforce the global context awareness and improve the accuracy of curve fitting. Firstly, a multi-scale feature extraction network is augmented by integrating a bottom-up feature propagation path and a non-local attention module, effectively fusing local details with long-range structural information to improve feature representation in complex scenarios. Secondly, a grouped parameter initialization strategy based on lane distribution priors is designed to encode geometric prior knowledge explicitly, optimizing model initialization and stabilizing lane position and orientation predictions. Additionally, a multi-task loss function combining polynomial coefficient prediction, lane point alignment, and average curvature constraints is introduced. This curvature-aware mechanism balances fitting accuracy for straight and curved lanes, mitigating data distribution bias. Furthermore, a label smoothing strategy based on polynomial fitting reduces annotation noise during training. Experiments demonstrate that the proposed method achieves F1 scores of 74.54% and 96.95% on the CULane and TuSimple datasets, respectively, while exhibiting superior robustness in challenging scenarios such as curves and occlusions. Ablation studies validate the contributions of each module, providing a high-precision end-to-end solution for lane detection tasks.</p>