Lane line detection at night is vulnerable to uneven illumination, strong light interference and image noise, resulting in blurred structure and unclear edges. To this end, this paper proposes DCDNet, a robust detection network that integrates frequency-domain decoupling and hierarchical attention. DCDNet consists of three major modules: LAFE adaptively adjusts the luminance response through the attention mechanism in the shallow layer to alleviate illumination interference; MSDCM models multi-scale contexts in the middle layer to enhance structural integrity. FHDM introduces decoupling of high and low frequencies in the frequency domain and gating fusion in the decoding stage, effectively separating details from background interference. The synergy of the three enhances the robustness against weakly textured lanes and pseudo-edge noise. On the CULane and TuSimple datasets, DCDNet significantly outperformed existing methods in Night scenarios, demonstrating good night-time adaptability and practical potential.