Abstract: The method based on semantic segmentation regards lane detection as a pixel segmentation problem, which is difficult to solve the efficiency problem and the challenging scenes such as severe occlusion and extreme lighting conditions. To solve these problems, a lane detection model based on multi-level feature fusion with location information enhancement is proposed to improve detection accuracy and address various challenging scenes. A segmentation module with multi-level feature fusion was used to extract more comprehensive lane position information. The position information of lane key points was fused with classification features to enhance the extraction of lane structure information and global context information. Compared with the best row-column classification model UFLDV2 on the CULane dataset, the F1 score is improved by 1.1 percentage points with a detection speed of 108FPS, which can meet the complex road conditions and real-time requirements in actual driving environments. It achieves considerable accuracy on the TuSimple dataset, and the lightweight version can even achieve over 190FPS.