Lane Graph as Path: Continuity-preserving Path-wise Modeling for Online Lane Graph Construction

TL;DR

The paper tackles online lane graph construction for autonomous driving, addressing the fragmentation of pixel-wise and piece-wise methods by introducing a path-centric formulation. LaneGAP jointly learns continuous directed paths from onboard sensors via Graph2Path, a set-based path detector, and Path2Graph to reconstruct a lane graph, using Polyline or Bezier path representations. It introduces graph-centric metrics (TOPO and Junction TOPO) and demonstrates superior accuracy and speed over baselines on nuScenes, Argoverse2, and OpenLane-V2, with ablations validating design choices. The work highlights the practical impact of path-oriented lane graphs for planning and as a foundational module for motion prediction and planning pipelines.

Abstract

Online lane graph construction is a promising but challenging task in autonomous driving. Previous methods usually model the lane graph at the pixel or piece level, and recover the lane graph by pixel-wise or piece-wise connection, which breaks down the continuity of the lane and results in suboptimal performance. Human drivers focus on and drive along the continuous and complete paths instead of considering lane pieces. Autonomous vehicles also require path-specific guidance from lane graph for trajectory planning. We argue that the path, which indicates the traffic flow, is the primitive of the lane graph. Motivated by this, we propose to model the lane graph in a novel path-wise manner, which well preserves the continuity of the lane and encodes traffic information for planning. We present a path-based online lane graph construction method, termed LaneGAP, which end-to-end learns the path and recovers the lane graph via a Path2Graph algorithm. We qualitatively and quantitatively demonstrate the superior accuracy and efficiency of LaneGAP over conventional pixel-based and piece-based methods on the challenging nuScenes and Argoverse2 datasets under controllable and fair conditions. Compared to the recent state-of-the-art piece-wise method TopoNet on the OpenLane-V2 dataset, LaneGAP still outperforms by 1.6 mIoU, further validating the effectiveness of path-wise modeling. Abundant visualizations in the supplementary material show LaneGAP can cope with diverse traffic conditions. Code is released at \url{https://github.com/hustvl/LaneGAP}.

Figures (7)

• Figure 1: Modeling comparison.(a) Pixel-wise modeling hdmapnet utilizes a predefined Graph2Pixel algorithm to rasterize the lane graph into a segmentation map and a direction map on dense BEV pixels, and heuristic Pixel2Graph post-processing is needed to recover the lane graph from the predicted segmentation map $V_{\text{pixel}}$ and direction map $D_{\text{pixel}}$ (direction map is not drawn here for simplicity). (b) Piece-wise modeling stsu utilizes a predefined Graph2Piece algorithm to split the lane graph into a set of pieces and the connectivity matrix among pieces, and then it merges the predicted pieces $\mathcal{V}_{\text{piece}}$ to the graph with the Piece2Graph algorithm based on predicted connectivity $E_{\text{piece}}$. (c) The proposed path-wise modeling translates the lane graph into complete paths with a predefined Graph2Path algorithm to traverse the graph. We perform path detection and adopt a Path2Graph algorithm to recover the lane graph.
• Figure 2: Qualitative comparison of pixel-wise HDMapNet, piece-wise MapTR, and path-wise LaneGAP on the nuScenes and Argoverse2 val splits. The top 2 rows are from the nuScenes val split, and the bottom 2 rows are from the Argoverse2 val split. Different colors indicate different instances. More qualitative comparisons are available in the supplementary.
• Figure 3: Overview of LaneGAP.
• Figure 4: Visualizations for ablations. The visualized ablation experiments are under a 24-epoch training schedule. The multi-modality method provides more accurate lane graph, and the Bezier method outputs smoother paths.
• Figure 5: Qualitative comparisons of increasing junction points. We qualitatively compare path-wise LaneGAP with pixel-wise HDMapNet and piece-wise MapTR on lane graphs with increasing junction points. All models perform well on simple lane graphs with few junction points. However, as the number of junction points increases, HDMapNet and MapTR struggle to predict reasonable lane graphs, while our proposed LaneGAP continues to deliver robust results.