The UFSA-LD algorithm faces challenges in extracting the thin and long structural features of lane lines, such as information loss, difficulty in capturing long-distance context, and insensitivity to boundary detail recognition. This paper proposes a lane line detection algorithm based on multi-scale atrous feature fusion attention: an MDFA module is added to the UFSA-LD auxiliary segmentation branch, and the receptive field of the network is expanded through atrous spatial pyramid pooling (ASPP) to capture lane features at multiple scales; a fusion channel and spatial attention mechanism (FCBAM) is used to filter out interfering information from channel and spatial dimensions, enhancing the representation of key features. The introduction of the Dice Loss loss function focuses more on the edges and local structural information of the lane lines. Experimental results show that the detection accuracy of the improved model on the TuSimple dataset has been increased from 95.81% to 96.03%; the F1 metric on the CULane dataset has improved by 1.8 compared to the original, validating the effectiveness of the model improvement.