Challenges in lane detection, such as distractors, illumination variations and cracks, necessitate a robust detection and extraction method based on road constraints. First, a large number of images for Dongfeng Park scene was collected to create a custom lane line dataset, including line and road area annotations. Second, we define a lane line segmentation loss function with road constraints to enhance the ENet-SAD model, yielding accurate segmentation predictions. Finally, we employ the dense CRF algorithm to refine the lane line probability map. Experimental results show that our model boosts the mean IoU and F1 score by 2.23% and 2.35% compared to ENet-SAD, especially in various complex scenarios including strong light, shadow, and occlusion. The proposed loss function can further enhance the performance of existing lane line detection models, including ENet-SAD and SCNN.