Radar pointcloud data pre-processing design#

Overview#

Pipeline#

This diagram describes the pre-process pipeline for radar pointcloud.

radar-pointcloud-sensing

Interface#

Input
- Radar data from device
- Twist information of ego vehicle motion
Output
- Dynamic radar pointcloud (ros-perception/radar_msgs/msg/RadarScan.msg)
- Noise filtered radar pointcloud (sensor_msgs/msg/Pointcloud2.msg)

Note#

In the sensing layer, the radar pre-process packages filter noise through the ros-perception/radar_msgs/msg/RadarScan.msg message type with sensor coordinate.
For use of radar pointcloud data by LiDAR packages, we would like to propose a converter for creating sensor_msgs/msg/Pointcloud2.msg from ros-perception/radar_msgs/msg/RadarScan.msg.

Reference implementations#

Data message for radars#

Autoware uses radar objects data type as radar_msgs/msg/RadarScan.msg. In detail, please see Data message for radars.

Device driver for radars#

Autoware support ros-perception/radar_msgs/msg/RadarScan.msg and autoware_auto_perception_msgs/msg/TrackedObjects.msg for Radar drivers.

In detail, please see Device driver for radars.

Basic noise filter#

radar_threshold_filter

This package removes pointcloud noise with low amplitude, edge angle, too near pointcloud by threshold. The noise depends on the radar devices and installation location.

Filter to static/dynamic pointcloud#

radar_static_pointcloud_filter

This package extracts static/dynamic radar pointcloud by using doppler velocity and ego motion. The static radar pointcloud can be used for localization like NDT scan matching, and the dynamic radar pointcloud can be used for dynamic object detection.

Message converter from RadarScan to Pointcloud2#

radar_scan_to_pointcloud2

For convenient use of radar pointcloud within existing LiDAR packages, we suggest a radar_scan_to_pointcloud2_convertor package for conversion from ros-perception/radar_msgs/msg/RadarScan.msg to sensor_msgs/msg/Pointcloud2.msg.

	LiDAR package	Radar package
message	`sensor_msgs/msg/Pointcloud2.msg`	`ros-perception/radar_msgs/msg/RadarScan.msg`
coordinate	(x, y, z)	(r, θ, φ)
value	intensity	amplitude, doppler velocity

For considered use cases,

Use pointcloud_preprocessor for radar scan.
Apply obstacle segmentation like ground segmentation to radar points for LiDAR-less (camera + radar) systems.

Appendix#

Discussion#

Radar architecture design is discussed as below.