Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Continuity Preserving Online CenterLine Graph Learning

Yunhui Han, Kun Yu, Zhiwei Li

TL;DR

This work tackles the challenge of constructing continuity-preserving centerline graphs for autonomous driving. It introduces CGNet, a DETR-inspired, end-to-end framework with three novel modules: Junction Aware Query (JAQ) for precise junction localization, Bézier Space Connection (BSC) to enforce smoothness in a Bézier space, and Iterative Topology Refinement (ITR) to iteratively refine connectivity. By modeling lanes as non-overlapped segments with topology and enforcing continuity through both point- and segment-level constraints, CGNet achieves state-of-the-art results on nuScenes and Argoverse2 across fine-grained point- and segment-level metrics. The approach promises practical benefits for online map reconstruction and downstream planning by producing continuous, topologically sound centerline graphs in real time, supported by comprehensive ablations and scalability analyses.

Abstract

Lane topology, which is usually modeled by a centerline graph, is essential for high-level autonomous driving. For a high-quality graph, both topology connectivity and spatial continuity of centerline segments are critical. However, most of existing approaches pay more attention to connectivity while neglect the continuity. Such kind of centerline graph usually cause problem to planning of autonomous driving. To overcome this problem, we present an end-to-end network, CGNet, with three key modules: 1)Junction Aware Query Enhancement module, which provides positional prior to accurately predict junction points; 2)Bézier Space Connection module, which enforces continuity constraints on any two topologically connected segments in a Bézier space; 3) Iterative Topology Refinement module, which is a graph-based network with memory to iteratively refine the predicted topological connectivity. CGNet achieves state-of-the-art performance on both nuScenes and Argoverse2 datasets.

Continuity Preserving Online CenterLine Graph Learning

TL;DR

This work tackles the challenge of constructing continuity-preserving centerline graphs for autonomous driving. It introduces CGNet, a DETR-inspired, end-to-end framework with three novel modules: Junction Aware Query (JAQ) for precise junction localization, Bézier Space Connection (BSC) to enforce smoothness in a Bézier space, and Iterative Topology Refinement (ITR) to iteratively refine connectivity. By modeling lanes as non-overlapped segments with topology and enforcing continuity through both point- and segment-level constraints, CGNet achieves state-of-the-art results on nuScenes and Argoverse2 across fine-grained point- and segment-level metrics. The approach promises practical benefits for online map reconstruction and downstream planning by producing continuous, topologically sound centerline graphs in real time, supported by comprehensive ablations and scalability analyses.

Abstract

Lane topology, which is usually modeled by a centerline graph, is essential for high-level autonomous driving. For a high-quality graph, both topology connectivity and spatial continuity of centerline segments are critical. However, most of existing approaches pay more attention to connectivity while neglect the continuity. Such kind of centerline graph usually cause problem to planning of autonomous driving. To overcome this problem, we present an end-to-end network, CGNet, with three key modules: 1)Junction Aware Query Enhancement module, which provides positional prior to accurately predict junction points; 2)Bézier Space Connection module, which enforces continuity constraints on any two topologically connected segments in a Bézier space; 3) Iterative Topology Refinement module, which is a graph-based network with memory to iteratively refine the predicted topological connectivity. CGNet achieves state-of-the-art performance on both nuScenes and Argoverse2 datasets.
Paper Structure (18 sections, 17 equations, 4 figures, 5 tables)

This paper contains 18 sections, 17 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Centerline Graph
  5. Preliminary on Bézier Curve.
  6. CGNet
  7. Architecture Overview.
  8. Junction Aware Query Enhancement Module.
  9. Bézier Space Connection Module.
  10. Iterative Topology Refinement Module.
  11. End-to-end Training
  12. Experiments
  13. Dataset
  14. Evaluation Metrics
  15. Implementation Details
  16. ...and 3 more sections

Figures (4)

  • Figure 1: The motivation. Top: A toy example which illustrates the centerline graph and the impact of overlooking the continuity. Bottom: Comparison with MapTR and TopoNet. They predicts inaccurate position of junction points and wrong topology, all leading to the discontinuous path. Our CGNet obtain the continuous path.
  • Figure 2: The overview architecture of CGNet. CGNet following the end-to-end paradigm of DETR, which takes 6 surrounding view images as inputs and output centerline graph without any post-processing. Three elaborately designed modules are utilized to preserving continuity, helping predict a continuous and smooth path.
  • Figure 3: The detailed illustration of our proposed modules. Left: Junction Aware Query Enhancement Module, which provides positional prior of junction to queries. Middle: Bézier Space Connection Module, which enforces continuity constraints on any two connected segments in the Bézier space. Right: Iterative Topology Refinement Module, which outputs connectivity of centerline segments in an iterative manner. "Memory Shortcut" means taking the output from previous layer.
  • Figure 4: Qualitative comparisons under different weather and lighting conditions on nuScenes. CGNet predicts more accurate position of junction points and correct topology, leading to a more continuous and smooth path compared to MapTR and TopoNet. CGNet demonstrats stronger robustness under different conditions.