LaneGraph2Seq: Lane Topology Extraction with Language Model Via Vertex-Edge Encoding and Connectivity Enhancement

Understanding road structures is crucial for autonomous driving. Intricateroad structures are often depicted using lane graphs, which include centerlinecurves and connections forming a Directed Acyclic Graph (DAG). Accurateextraction of lane graphs relies on precisely estimating vertex and edgeinformation within the DAG. Recent research highlights Transformer-basedlanguage models' impressive sequence prediction abilities, making themeffective for learning graph representations when graph data are encoded assequences. However, existing studies focus mainly on modeling verticesexplicitly, leaving edge information simply embedded in the network.Consequently, these approaches fall short in the task of lane graph extraction.To address this, we introduce LaneGraph2Seq, a novel approach for lane graphextraction. It leverages a language model with vertex-edge encoding andconnectivity enhancement. Our serialization strategy includes a vertex-centricdepth-first traversal and a concise edge-based partition sequence.Additionally, we use classifier-free guidance combined with nucleus sampling toimprove lane connectivity. We validate our method on prominent datasets,nuScenes and Argoverse 2, showcasing consistent and compelling results. OurLaneGraph2Seq approach demonstrates superior performance compared tostate-of-the-art techniques in lane graph extraction.