TLDR: This paper introduces a custom, open-source workflow for creating high-definition maps for autonomous vehicle simulations, specifically for AWSIM and Autoware. It addresses the challenges of complex existing methods by using OpenStreetMap data and lightweight tools to generate 3D mesh, point cloud, and Lanelet2 files, enabling efficient testing of autonomous systems in simulated environments. The workflow was successfully demonstrated by creating a parking lot map for Ontario Tech University, proving its effectiveness for early-stage development and rapid prototyping.
The rapid advancement of autonomous vehicle (AV) technology has brought with it a critical need for extensive and rigorous simulation testing. These simulations rely heavily on accurate and adaptable maps, which serve as the backbone for everything from vehicle localization to path planning and scenario testing. However, creating these simulation-ready maps can often be a complex and resource-intensive endeavor, especially when dealing with sophisticated simulators like CARLA or AWSIM.
Many existing map creation workflows demand significant computational power or are tightly coupled to specific simulators, which limits flexibility for developers. Recognizing this challenge, researchers Zubair Islam, Ahmaad Ansari, George Daoud, and Mohamed El-Darieby from Ontario Tech University have developed a custom workflow designed to simplify and streamline the map creation process for AV development. Their work, detailed in their paper A Workflow for Map Creation in Autonomous Vehicle Simulations, demonstrates this through the generation of a 3D map of a parking lot at their university.
Addressing the Simulation Gap
The motivation for this research stemmed from a practical limitation: at the time of development, popular AV simulation platforms like AWSIM and Autoware offered only a single, large city environment map. Crucially, this map lacked a parking lot, an essential feature for testing real-world AV deployment in low-speed, complex scenarios involving interactions with other vehicles. While documentation for creating custom environments existed, it was often difficult to interpret and follow, highlighting the need for a more straightforward solution.
The core requirement for a custom environment in AWSIM involves three key files: a Lanelet2 OSM file, a Point Cloud Data (PCD) file, and a 3D mesh file. The team’s custom workflow leverages multiple open-source tools to generate these necessary files efficiently. By starting with an OpenStreetMap (OSM) file, the workflow enables the creation of custom environments for any outdoor area available on OSM, which is a widely used open-source geospatial data resource.
The Custom Workflow Explained
The proposed workflow consists of four main steps, each utilizing specific tools to transform raw geospatial data into a functional simulation map:
- OpenStreetMap (OSM) Selection: Users begin by selecting a desired real-world location from the OpenStreetMap website and exporting it as an OSM file. This file contains detailed geographical information, including nodes, ways, relations, and tags that define the area’s features.
- Automated Mapping Pipeline Docker Container: This container integrates several tools for automated processing. First, OSM2World converts the OSM file into a 3D mesh, generating OBJ, MTL, and PNG files that represent the three-dimensional model of the location. Next, CloudCompare is used to import this 3D mesh and extract a point cloud, which is a collection of data points in a 3D coordinate system representing the shape of the mesh. Finally, the Point Cloud Library (PCL) processes this point cloud, correcting its orientation to a top-down view and converting it to a binary format, making it ready for Autoware.
- Vector Map Builder: This tool, provided by Tier IV, is crucial for creating a Lanelet2 vector map. Lanelet2 is a specialized format for AV simulations that defines road networks, lanes, and other essential road features. Users can import the processed point cloud file and manually define lanes, parking lots, and parking spaces, which are then exported as a Lanelet2 OSM file.
- Python Script for OSM Manipulation: A custom Python script is used to nullify all latitude and longitude fields within the generated Lanelet2 OSM file. This step is vital for ensuring proper functionality with the Autoware software, as un-nullified coordinates can cause lanes to malfunction in the simulation.
Integration and Results
Once these files are generated and correctly named (e.g., lanelet2_map.osm and pointcloud_map.pcd), they can be seamlessly imported into Autoware and AWSIM. Autoware requires the Lanelet2 OSM and PCD files, while AWSIM uses the 3D mesh and Lanelet2 OSM file, with additional steps to enable mesh colliders for interaction within the simulation environment.
The researchers successfully implemented this workflow to create a functional 3D map of Ontario Tech University’s SIRC parking lot. This map was then tested in both Autoware and AWSIM. The ego vehicle (the autonomous vehicle being simulated) was correctly spawned and localized within the environment, with all its sensors functioning. After setting a goal pose within a parking spot, both simulators accurately mimicked each other, with the vehicle successfully reaching its destination. This demonstrated the map’s effective integration and the ego vehicle’s proper localization and navigation capabilities.
Also Read:
- HDSim: Crafting Human-like Traffic for Advanced Self-Driving Tests
- Drone Navigation in 3D Urban Environments: A Comparative Study of Path Planning Algorithms
Future Outlook
While the 3D models generated by OSM2World may have limitations in geometric accuracy compared to real-world scans (for instance, a soccer dome might appear as a simple rectangular building), the workflow offers significant practical advantages. It operates entirely offline using publicly available OpenStreetMap data and lightweight, open-source tools, making it highly accessible and easy to deploy, especially for early-stage development, academic research, and rapid prototyping where real-world precision isn’t the primary concern.
The workflow’s flexibility means it can be adapted for use with other simulators, though this was not the focus of the current paper. Future work aims to improve model accuracy by incorporating SLAM (Simultaneous Localization and Mapping) technologies, optimize the workflow for broader simulator compatibility, and explore more flexible handling of latitude and longitude values to enhance map generation accuracy.
This custom workflow provides a valuable solution for creating high-definition maps for autonomous vehicle simulations, addressing a critical need for diverse and adaptable testing environments. By simplifying the map creation process, it empowers developers to accelerate the testing and validation of AV technologies in a controlled and cost-effective manner.
