Perception launch files#

Overview#

The Autoware perception stack starts with autoware.launch.xml, as the Launch Autoware page explains. If launch_perception is true, autoware.launch.xml includes two component launch files of the autoware_launch package: component_perception.launch.xml and component_traffic_light.launch.xml. These files include the launch files of the autoware_perception_launch package. The following diagram shows the perception launch flow in the autoware_launch repository.

component_perception.launch.xml includes the LiDAR perception pipeline, and component_traffic_light.launch.xml includes one traffic light recognition pipeline.

Note

The Autoware project is a large project. Therefore, as we manage the Autoware project, we utilize specific arguments in the launch files. ROS 2 offers an argument-overriding feature for these launch files. Please refer to the official ROS 2 launch documentation for further information. For instance, if we define an argument at the top-level launch, it will override the value on lower-level launches.

component_perception.launch.xml#

The component_perception.launch.xml launch file is the main perception launch file of the autoware_launch package. It starts the LiDAR pipeline in the /perception namespace: the common point cloud downsampling (optional), obstacle segmentation, the occupancy grid map, and object recognition. There is no single perception.launch.xml file any more. The parameter files are in the autoware_perception_config package. You can change the perception launch arguments in component_perception.launch.xml. These are some of its arguments:

occupancy_grid_map_method: This argument selects the method of the occupancy grid map. For more information, see the probabilistic_occupancy_grid_map package. The options are pointcloud_based (the default), laserscan_based, and multi_lidar_pointcloud_based. To change it to laserscan_based, change it here:

- <arg name="occupancy_grid_map_method" default="pointcloud_based" description="options: pointcloud_based, laserscan_based, multi_lidar_pointcloud_based"/>
+ <arg name="occupancy_grid_map_method" default="laserscan_based" description="options: pointcloud_based, laserscan_based, multi_lidar_pointcloud_based"/>

detected_objects_filter_method: This argument selects the filter method for the detected objects. For more information about the lanelet filter and the position filter, see the detected_object_validation package. The default filter method is lanelet_filter. To change it to position_filter, change it here:
```
- <arg name="detected_objects_filter_method" default="lanelet_filter" description="options: lanelet_filter, position_filter"/>
+ <arg name="detected_objects_filter_method" default="position_filter" description="options: lanelet_filter, position_filter"/>
```

detected_objects_validation_method: This argument selects the validation method for the detected objects. For more information about the validation methods, see the detected_object_validation package. The validator runs only if use_object_validator is true. Its default value is false. The default validation method is obstacle_pointcloud. To change it to occupancy_grid, enable the validator and use laserscan_based as occupancy_grid_map_method:

- <arg name="use_object_validator" default="false" description="use obstacle_pointcloud based object validator"/>
+ <arg name="use_object_validator" default="true" description="use obstacle_pointcloud based object validator"/>
- <arg name="occupancy_grid_map_method" default="pointcloud_based" description="options: pointcloud_based, laserscan_based, multi_lidar_pointcloud_based"/>
+ <arg name="occupancy_grid_map_method" default="laserscan_based" description="options: pointcloud_based, laserscan_based, multi_lidar_pointcloud_based"/>
- <arg name="detected_objects_validation_method" default="obstacle_pointcloud" description="options: obstacle_pointcloud, occupancy_grid (must be used with laserscan_based_occupancy_grid_map)"/>
+ <arg name="detected_objects_validation_method" default="occupancy_grid" description="options: obstacle_pointcloud, occupancy_grid (must be used with laserscan_based_occupancy_grid_map)"/>

lidar_detection_model: This argument selects the LiDAR detection model. The default value is centerpoint. For the other models, see the Perception component reference implementation.

use_cuda_ground_segmentation and use_semantic_segmentation_ptv3: These arguments select another ground segmentation backend: the CUDA version of the rule-based filter, or the PTv3 semantic segmentation. The default value of both is false. If use_semantic_segmentation_ptv3 is true, object recognition uses object_recognition_lidar_ptv3.launch.xml instead of object_recognition_lidar.launch.xml. Then detected_objects_filter_method, detected_objects_validation_method, and use_object_validator have no effect.

Note

You can also use these arguments as command-line arguments:

ros2 launch autoware_launch autoware.launch.xml ... detected_objects_filter_method:=lanelet_filter occupancy_grid_map_method:=laserscan_based ...

component_perception.launch.xml also launches the perception analytics publisher (use_perception_analytics_publisher, true by default) and the perception online evaluator (use_perception_online_evaluator, false by default) of autoware_perception_online_evaluator.

Use cameras and radars#

Object recognition in autoware_launch uses only LiDARs. The sensor kits have no radars, and object recognition does not use their cameras. So component_perception.launch.xml uses object_recognition_lidar.launch.xml, and Autoware 2.0.0 has no perception_mode argument. If your vehicle has cameras or radars, include another object recognition launch file of autoware_perception_launch in your fork of component_perception.launch.xml, for example object_recognition_camera_lidar_radar.launch.xml. For the list of these files and their detectors, see the Perception component reference implementation.

These launch files have arguments for the camera topics. Give them the topics of your cameras, for example:

<include file="$(find-pkg-share autoware_perception_launch)/launch/object_recognition/object_recognition_camera_lidar_radar.launch.xml">
  <arg name="number_of_cameras" value="1"/>
  <arg name="input/camera0/image" value="/sensing/camera/camera0/image_rect_color"/>
  <arg name="input/camera0/info" value="/sensing/camera/camera0/camera_info"/>
  <arg name="input/camera0/rois" value="/perception/object_recognition/detection/rois0"/>
  ...
</include>

The parameters of the camera-LiDAR fusion nodes are in autoware_perception_config. For example, remove_unknown of roi_cluster_fusion is in config/object_recognition/detection/image_projection_based_fusion/roi_cluster_fusion.param.yaml. Its value is false.

component_traffic_light.launch.xml#

The component_traffic_light.launch.xml launch file starts traffic light recognition in the /perception/traffic_light_recognition namespace, if use_traffic_light_recognition is true (the default). It includes one of three pipelines of autoware_perception_launch:

Arguments	Launch file
`traffic_light_recognition/use_high_accuracy_detection:=true` and `high_accuracy_detection_type:=fine_detection` (the defaults)	`traffic_light_recognition_fine.launch.xml`
`traffic_light_recognition/use_high_accuracy_detection:=true` and `high_accuracy_detection_type:=whole_image_detection`	`traffic_light_recognition_whole_image.launch.xml`
`traffic_light_recognition/use_high_accuracy_detection:=false`	`traffic_light_recognition_rough.launch.xml`

The full name of the second argument is traffic_light_recognition/high_accuracy_detection_type. Each pipeline includes a detector launch file and traffic_light_fusion.launch.xml. The fine and the rough pipelines also include traffic_light_occlusion_predictor.launch.xml. The camera topics are fixed inside each pipeline launch file.

You can add other arguments to component_perception.launch.xml and component_traffic_light.launch.xml like these examples.