Map component design#

Overview#

Autoware uses high-definition maps of the driving environment for localization, route planning, traffic light detection, and the prediction of the paths of pedestrians and other vehicles.

This page describes the map component of Autoware: its requirements, architecture, interfaces, and the formats of the map files.

Requirements#

The map component gives two types of information to the rest of the stack:

• Semantic information about the roads, as a vector map.
• Geometric information about the environment, as a point cloud map. This map is optional.

A vector map contains accurate information about the road network, the lane geometry, and the traffic lights. Route planning, traffic light detection, and the prediction of the paths of other vehicles and pedestrians need the vector map.

Autoware uses the 3D point cloud map for LiDAR-based localization and for some perception functions. To find the position and orientation of the vehicle, the localization matches a live scan from one or more LiDARs against the point cloud map. As a result, the accuracy of the point cloud map controls the accuracy of the localization. If the vehicle has a different localization method with enough accuracy, for example camera-based localization, Autoware can run without a point cloud map.

Autoware also needs a projection information file. This file defines how the local map coordinates relate to the geodetic coordinates (latitude, longitude, and altitude).

Architecture#

The map component runs these nodes in the /map namespace:

All the packages are in Autoware Core, except autoware_map_tf_generator, which is in Autoware Universe.

The launch file component_map.launch.xml of autoware_launch starts the map component. It includes one launch file of the autoware_map_launch package for each feature. The arguments use_pointcloud_map, use_lanelet2_map, use_map_projection, use_map_tf_generator, and use_map_hash_generator turn each feature on or off. All of them are true by default.

The launch file reads the map files from the directory in the map_path argument:

map_path
├── lanelet2_map.osm               # the vector map (the lanelet2_map_file argument)
├── pointcloud_map.pcd             # the point cloud map (the pointcloud_map_file argument), a file or a directory
├── pointcloud_map_metadata.yaml   # needed for a divided point cloud map
├── lanelet2_map_metadata.yaml     # optional, for a divided vector map
└── map_projector_info.yaml        # the projection information

Component interface#

Input to the map component#

The map component reads only files:

• The point cloud map and its metadata file
• The vector map
• The projection information

Output from the map component#

Map specification#

Point cloud map#

The point cloud map must meet these requirements:

• The point cloud map must use the same projection as the one that map_projection_loader publishes. Then it agrees with the Lanelet2 map and with the packages that convert between local and geodetic coordinates. For more information, see the README of autoware_map_projection_loader.
• It must be in the PCD (Point Cloud Data) file format. It can be one PCD file or many PCD files (a divided map).
• Each point in the map must contain X, Y, and Z coordinates.
• Each point can also contain an intensity or an RGB value.
• It must cover the entire operational area of the vehicle. Add a buffer zone around this area that matches the detection range of the sensors on the vehicle.
• The design recommends a point spacing of about 0.1 m to 0.2 m. The point cloud map pre-processing tutorial downsamples the map with a voxel size in this range to decrease the computation cost and the memory bandwidth load. This is a recommendation, not a check that Autoware does. The NDT scan matcher makes its own voxel grid from the map, with a resolution of 2.0 m by default.
• It can be in local or global coordinates. To use GNSS data for localization, it must be in global coordinates (georeferenced).

For a large map, use a divided map. The point cloud map loader can then send only the parts of the map near the vehicle (partial and differential loading). A divided map must be cut along lines parallel to the x-axis and the y-axis into cells of the same size, about 20 m by 20 m. It also needs a metadata file. autoware_pointcloud_divider can divide a map and write the metadata file. For the details of the divided map format, see the README of autoware_map_loader.

Vector map#

The vector map must meet these requirements:

• It must be in Lanelet2 format, with the additional rules of Autoware.
• It must contain the shape and position of the lanes, traffic lights, stop lines, crosswalks, parking spaces, and parking lots.
• Except at the start or end of a road, each lanelet must have correct connections to its predecessor, successors, left neighbor, and right neighbor.
• Each lanelet must contain its traffic rules: the speed limit, the right of way, the traffic direction, and the related traffic lights, stop lines, and traffic signs.
• It must cover the entire operational area of the vehicle.

The vector map can be one .osm file or a directory of .osm files. The Lanelet2 map loader merges all the files of a directory into one map. A map with more than one file needs a lanelet2_map_metadata.yaml file.

For the detailed rules to create a vector map, see the vector map creation requirement specifications.

Projection information#

The projection information is a YAML file, map_projector_info.yaml, that map_projection_loader reads. It contains these fields:

For more information, see the README of autoware_map_projection_loader.