Procedural Generation of High-Definition Road Networks for Autonomous Vehicle Testing and Traffic Simulations
- Golam Md Muktadir ,
- Abdul Jawad - University of California Santa Cruz, Computational Media Department, USA ,
- Ishaan Paranjape - University of California Santa Cruz, Computational Media Department, USA ,
- Jim Whitehead - University of California Santa Cruz, Computational Media Department, USA ,
- Aleksey Shepelev
ISSN: 2574-0741, e-ISSN: 2574-075X
Published May 10, 2022 by SAE International in United States
Citation: Muktadir, G., Jawad, A., Paranjape, I., Whitehead, J. et al., "Procedural Generation of High-Definition Road Networks for Autonomous Vehicle Testing and Traffic Simulations," SAE Intl. J CAV 6(1):2023, https://doi.org/10.4271/12-06-01-0007.
Effective simulation-based testing of autonomous vehicles requires the exploration of vehicle performance against a wide variety of rare and unusual road and intersection geometries. We present a high-definition road and intersection generator called JunctionArt. JunctionArt takes as input a series of control lines and generates a series of roads and intersections which conforms to them. Roads exhibit different types of lane types such as turn lanes, one-way streets, and multiple lanes, while intersections feature a range of incident roads (three to seven incident roads), leading to a variety of geometries and interior connecting lanes. These roads are output in the OpenDRIVE format and, hence, are interoperable with a wide range of tools and simulation environments. Multiple metrics are computed over generated roads—field of view (FOV), maximum turn curvature (maxCurvature), corner deviation angle (cornerDeviation), complexity, conflictArea, and the number of interior connection lanes—and are used to perform an expressive range analysis. This analysis finds that JunctionArt is capable of creating rare and unusual intersection situations, i.e., representatives of the long tail of infrequent road configurations. Interoperability with third-party simulation tools and environments RoadRunner, Carla, and esmini is demonstrated.