Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LGmap: Local-to-Global Mapping Network for Online Long-Range Vectorized HD Map Construction

Kuang Wu, Sulei Nian, Can Shen, Chuan Yang, Zhanbin Li

TL;DR

This work tackles online construction of long-range, vectorized HD maps for mapless autonomous driving. It introduces LGmap, which fuses forward and backward view projections via Symmetric View Transformation (SVT) and aggregates temporal information through Hierarchical Temporal Fusion (HTF) to build stable long-range maps; a simplified ped-crossing representation accelerates decoding. The paper also details specialized components for area handling, lane segments, traffic elements, and lane topology (TopoMLP-based LLT and LTT), and demonstrates significant performance gains on OpenLaneV2 with a final UniScore of 0.66. Overall, LGmap advances real-time, online HD map construction by integrating robust visual-LiDAR cues, temporal fusion, and topology reasoning to support long-range navigation in dynamic urban environments.

Abstract

This report introduces the first-place winning solution for the Autonomous Grand Challenge 2024 - Mapless Driving. In this report, we introduce a novel online mapping pipeline LGmap, which adept at long-range temporal model. Firstly, we propose symmetric view transformation(SVT), a hybrid view transformation module. Our approach overcomes the limitations of forward sparse feature representation and utilizing depth perception and SD prior information. Secondly, we propose hierarchical temporal fusion(HTF) module. It employs temporal information from local to global, which empowers the construction of long-range HD map with high stability. Lastly, we propose a novel ped-crossing resampling. The simplified ped crossing representation accelerates the instance attention based decoder convergence performance. Our method achieves 0.66 UniScore in the Mapless Driving OpenLaneV2 test set.

LGmap: Local-to-Global Mapping Network for Online Long-Range Vectorized HD Map Construction

TL;DR

This work tackles online construction of long-range, vectorized HD maps for mapless autonomous driving. It introduces LGmap, which fuses forward and backward view projections via Symmetric View Transformation (SVT) and aggregates temporal information through Hierarchical Temporal Fusion (HTF) to build stable long-range maps; a simplified ped-crossing representation accelerates decoding. The paper also details specialized components for area handling, lane segments, traffic elements, and lane topology (TopoMLP-based LLT and LTT), and demonstrates significant performance gains on OpenLaneV2 with a final UniScore of 0.66. Overall, LGmap advances real-time, online HD map construction by integrating robust visual-LiDAR cues, temporal fusion, and topology reasoning to support long-range navigation in dynamic urban environments.

Abstract

This report introduces the first-place winning solution for the Autonomous Grand Challenge 2024 - Mapless Driving. In this report, we introduce a novel online mapping pipeline LGmap, which adept at long-range temporal model. Firstly, we propose symmetric view transformation(SVT), a hybrid view transformation module. Our approach overcomes the limitations of forward sparse feature representation and utilizing depth perception and SD prior information. Secondly, we propose hierarchical temporal fusion(HTF) module. It employs temporal information from local to global, which empowers the construction of long-range HD map with high stability. Lastly, we propose a novel ped-crossing resampling. The simplified ped crossing representation accelerates the instance attention based decoder convergence performance. Our method achieves 0.66 UniScore in the Mapless Driving OpenLaneV2 test set.
Paper Structure (21 sections, 3 figures, 5 tables)

This paper contains 21 sections, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Pipeline
  4. Encoder
  5. Decoder
  6. Temporal fusion
  7. Loss functions
  8. Area
  9. Lane segments
  10. Traffic elements
  11. Lane-Lane topology
  12. Lane-Traffic topology
  13. Experiments
  14. Implementation details
  15. Ablation Study
  16. ...and 6 more sections

Figures (3)

  • Figure 1: The overall model architecture of LGmap. The entire model is consists of mainly six components: a image backbone equipped with SVT(Symmetric View Transformation), a hierarchical temporal fusion(HTF) module, a unified instance detection and segmentation predictor, a traffic elements detector(YOLO wang2024yolov9), a Lane-Lane Topology(LLT) and a Lane-TE Topology(LTT).
  • Figure 2: Stacking strategy and streaming strategy are same as StreamMapNet's yuan2024streammapnet summary. In order to demonstrate the effectiveness of long-range stacking for streaming-stacking strategy in figure, the stacking previous frame interval parameter is set to 2. Stacking strategy only fuses one previous frame in this figure, and actually it may fuse more than one frame.
  • Figure 3: The ped crossing form of MapTR, MachMap and LGmap.