Lane detection is a core task in autonomous driving and driver assistance systems. To address the issues of feature degradation and geometric mismatch in CLRNet in complex scenarios, this paper proposes a dual improvement strategy: embedding a Hierarchical Attention Network with Convolutional Block Attention Module (HAN-CBAM) to enhance the feature selection capability of the backbone network, and designing a noise-robust Dynamic Line IoU (DL-IoU) loss to improve the geometric matching accuracy of curved lanes. Experiments on a 1/10 scale subset of the CULane dataset demonstrate that the improved model achieves an F1@50 score of 79.8 (1.5 higher than the baseline) while maintaining real-time inference speed (171 FPS on RTX 3080 Ti). The attention mechanism significantly enhances detection continuity in occluded scenarios, and the dynamic loss function improves the stability of curved lane IoU, demonstrating superior robustness in complex environments such as night and crowded conditions. The balance between the comprehensive performance metric mF1 (55.46) and high-speed inference validates the practical value of this solution in autonomous driving systems.