Automatic Construction of Lane-level HD Maps for Urban Scenes

Yiyang Zhou, Yuichi Takeda, Masayoshi Tomizuka, Wei Zhan

Abstract

High definition (HD) maps have demonstrated their essential roles in enabling full autonomy, especially in complex urban scenarios. As a crucial layer of the HD map, lane-level maps are particularly useful: they contain geometrical and topological information for both lanes and intersections. However, large scale construction of HD maps is limited by tedious human labeling and high maintenance costs, especially for urban scenarios with complicated road structures and irregular markings. This paper proposes an approach based on semantic-particle filter to tackle the automatic lane-level mapping problem in urban scenes. The map skeleton is firstly structured as a directed cyclic graph from online mapping database OpenStreetMap. Our proposed method then performs semantic segmentation on 2D front-view images from ego vehicles and explores the lane semantics on a birds-eye-view domain with true topographical projection. Exploiting OpenStreetMap, we further infer lane topology and reference trajectory at intersections with the aforementioned lane semantics. The proposed algorithm has been tested in densely urbanized areas, and the results demonstrate accurate and robust reconstruction of the lane-level HD map.

Introduction

For autonomous vehicles to navigate safely through dense urban environments, they require high-definition (HD) maps that provide lane-level detail. Unlike standard navigation maps, HD maps specify exact lane boundaries, centerlines, and the complex connectivity of intersections. While highway mapping has reached a level of maturity due to standardized markings and simple geometries, urban environments remain a significant challenge. Frequent lane splits, "invisible" lanes within intersections, and irregular road features like bike lanes or bus stops make automated mapping difficult.

Current industry standards often rely on labor-intensive manual labeling or struggle with the non-planar nature of urban roads—such as the steep hills found in San Francisco. This research introduces an automated pipeline designed to bridge this gap, combining semantic scene understanding with a novel particle-filter exploration strategy and OpenStreetMap (OSM) priors to generate precise, structured road networks.

Methodology: From Pixels to Topology

The proposed system transforms raw vehicle sensor data into a structured Directed Cyclic Graph (DAG) through a multi-stage process.

Semantic Segmentation and BEV Projection

The pipeline begins with Semantic Scene Understanding. Using a DeepLab-v3+ architecture, the system segments 2D camera images into ten distinct classes, including ego-lanes, neighboring lanes, and various road markings.

A critical innovation in this work is the Topographical Bird’s-Eye-View (BEV) Projection. Most existing systems assume a flat ground plane, which leads to significant geometric distortions on hills or uneven pavement. By incorporating optional LiDAR data to create a 3D topographical mesh via Delaunay Triangulation, the system projects semantics onto the "true" ground shape. This ensures that lane widths and crosswalk geometries remain accurate regardless of the terrain.

The pipeline illustrates the transition from raw front-view images to semantic segmentation, topographical BEV projection, and the final lane-level map reconstruction.

Semantic-Particle Filter Exploration

To extract structured lanes from the semantic BEV map, the authors employ a Semantic-Particle Filter. Instead of simple line detection, "particles" (representing simulated vehicle paths) are released to explore the drivable areas. Particles that remain within valid lane semantics survive, while those that cross boundaries are terminated. This approach is inherently robust to faded or broken lane markings, as it focuses on the overall drivability of the scene.

Lane Regression and Intersection Inference

Surviving particle paths are clustered using DBSCAN to identify individual lanes. These paths are then smoothed into differentiable curves using natural splines. To handle intersections where markings often disappear, the system utilizes OpenStreetMap (OSM) as a topological skeleton. By understanding the coarse connectivity provided by OSM, the algorithm uses second-order Bezier curves to mathematically link incoming and outgoing lanes, ensuring a continuous and logically consistent map.

Topological Mapping and OSM Priors

The integration of OSM allows the system to move beyond disconnected lane segments toward a complete network. The physical road structures are mapped to a Directed Cyclic Graph (DAG), where nodes represent intersections and edges represent the lane-level trajectories.

The system integrates physical lane data with OpenStreetMap priors to construct a structured Directed Cyclic Graph (DAG) of the urban road network.

Experimental Results

The system was rigorously tested in two challenging urban environments: North Beach in San Francisco and Downtown Berkeley. These areas were chosen for their steep slopes, irregular multi-way intersections, and high density of road features.

  • Accuracy: In Downtown Berkeley, the system achieved a Root Mean Square (RMS) error of 0.24m for lane trajectories. Even in the challenging terrain of San Francisco, the RMS error remained impressively low at 0.33m.
  • Robustness: The topological inference module reached over 90% precision in correctly identifying lane connections within intersections without human intervention.
  • Topographical Impact: The authors demonstrated that using a LiDAR-based mesh for BEV projection significantly reduced mapping errors compared to "flat earth" assumptions, which is vital for achieving decimeter-level precision.

Conclusion and Future Work

This research presents a robust framework for the automatic construction of lane-level HD maps. By combining deep learning for scene understanding with classical filtering for exploration and OSM for global topology, the system handles the complexities of urban driving environments effectively.

The primary remaining challenge is occlusion; in heavy traffic, other vehicles can block the camera's view of essential road markings. Future work will explore "amodal prediction"—training models to infer hidden road structures—and performing semantic processing directly in the BEV domain to further improve reliability in congested city centers.

Summary updated

Abstract

High definition (HD) maps have demonstrated their essential roles in enabling full autonomy, especially in complex urban scenarios. As a crucial layer of the HD map, lane-level maps are particularly useful: they contain geometrical and topological information for both lanes and intersections. However, large scale construction of HD maps is limited by tedious human labeling and high maintenance costs, especially for urban scenarios with complicated road structures and irregular markings. This paper proposes an approach based on semantic-particle filter to tackle the automatic lane-level mapping problem in urban scenes. The map skeleton is firstly structured as a directed cyclic graph from online mapping database OpenStreetMap. Our proposed method then performs semantic segmentation on 2D front-view images from ego vehicles and explores the lane semantics on a birds-eye-view domain with true topographical projection. Exploiting OpenStreetMap, we further infer lane topology and reference trajectory at intersections with the aforementioned lane semantics. The proposed algorithm has been tested in densely urbanized areas, and the results demonstrate accurate and robust reconstruction of the lane-level HD map.
BibTeX
@inproceedings{zhou2021automatic,
  author = {Yiyang Zhou and Yuichi Takeda and Masayoshi Tomizuka and Wei Zhan},
  title = {{Automatic Construction of Lane-level HD Maps for Urban Scenes}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2021},
  url = {https://openreview.net/forum?id=17-XfVlmPy},
}
Figure