关键词: 深度学习；车道线检测；特征关联；相对起始点；双线；分岔线；聚类

Abstract: At present, lane line detection methods based on deep learning perform well in most scenarios, but they perform poorly in distinguishing lane line instances with complex topological structures. To solve such problems, the article proposes a method for expressing lane instances based on the extraction of relative starting points through dynamic feature association. Firstly, a deep learning model is utilized to extract the foreground points representing the lane line features. Then, based on the global feature information, the association relationship between the foreground points and the starting points of the lane lines is established. Finally, a lane instance clustering algorithm decoder based on dynamically adjusting the starting points is constructed to decode the lane features based on the relative starting points. Meanwhile, the Pull-Push loss is innovatively adopted to dynamically adjust the association state between the foreground point features and the starting point features. This method enables the foreground points to be associated at the relative starting point rather than the absolute starting point, which can effectively overcome the problem of lane line instance detection where the starting point is relatively close. This research has significantly improved the effect in dealing with the problem of lane line instance distinction under complex topologies.

Key words: deep learning; lane line detection; feature association; relative starting point; double lines; bifurcation line; clustering