This paper introduces a lane detection method that leverages a large-kernel attention module for improved global information capture in complex scenarios. The large-kernel attention mechanism enhances the model’s ability to capture global information by expanding the receptive field, thus improving performance. The main branch employs a lightweight residual network for feature extraction, introduces deformable convolution in some residual layers to enhance the model’s feature extraction capabilities, and combines with the large-kernel attention module for context information fusion, effectively addressing issues such as lane line occlusion and improving detection accuracy. An auxiliary segmentation branch captures long-distance spatial information correlations through a spatial attention module, performing pixel-by-pixel semantic segmentation of images, with the generated results used to assist network training. Tested on Tusimple and CuLane datasets, the approach outperforms traditional YOLO and Res methods, improving accuracy by 15.91% and 3.78%, respectively, on TuSimple, and showing robustness across various scenes in CULane.