Abstract
Lane detection is a crucial task in autonomous driving, but existing methods still struggle with occluded, missing, and complex road structures of lanes. To address these issues, we propose a graph attention-based 3D lane detection method-Graph-RMNet. This network enhances the model's understanding of the three-dimensional distribution of lanes by parsing the topological structure and semantic category relationships between them. First, we introduce a 3D positional query generation strategy that combines 3D spatial information with 2D image features to initialize queries, thereby enhancing the queries' perception of the three-dimensional distribution of lanes. Furthermore, we design a dual-path relation module that captures complex interactions through graph attention mechanisms and models lane relations from both spatial and categorical dimensions via a specially designed adaptive graph structure, improving the model's comprehension and expression of complex inter-lane relations. Ultimately, the prediction module outputs the 3D coordinates and categories of the lanes to construct a complete distribution of the lanes in 3D space. The experimental results show that Graph-RMNet achieves F-Score of 63.2% on the OpenLane dataset and F-Score of 80.63% on the ONCE-3DLanes dataset, both outperforming current popular algorithms. Notably, it demonstrates significant robustness in scenarios where lanes are occluded or missing.