A Novel Structure Prior-based Loss Function for Lane Detection

In the field of autonomous driving, lane detection is of fundamental importance in the trajectory planning decision and the subsequent lane departure. However, lane detection is still a challenging task due to many existing problems, such as lane defect and occlusion, the variation of the number of lanes, diverse lane types and real-time detection requirement. As a major problem, lane defect and occlusion can lead to inaccurate lane segmentation and thus produce wrong parameters for lane curve fitting. To solve this issue, this paper proposes a novel loss function based on the structure prior of lanes. More specifically, single lanes should always be spatially continuous and lanes of the same scene are likely to be parallel with each other. The proposed loss function makes the lane detection network more robust to segmentation outliers during fitting lane curves, and reduces fitting errors. Our loss function is conceptually simple and can be easily applied to many lane detection algorithms. As a demonstration, we replaced the loss function of one state-of-the-art lane detection model and retrained it on the same dataset. The experimental results confirm that the improved loss function can further improve the model performance and decrease the area error by up to 15.4%.