Open kubs0ne opened 3 years ago
Hello,
The reference point cloud can be given in a file or a ros point cloud message: https://github.com/carlosmccosta/dynamic_robot_localization#reference-map https://github.com/carlosmccosta/dynamic_robot_localization/blob/noetic-devel/yaml/schema/drl_configs.yaml#L189 https://github.com/carlosmccosta/dynamic_robot_localization/blob/noetic-devel/yaml/schema/drl_configs.yaml#L73
For 3 DoF localization you can specify the file in the parameter reference_pointcloud_filename
in the dynamic_robot_localization_system.launch.
Have a nice day,
Thank you very much! So that means I could specify the poitcloud filename, and then specify the topic of the live pointclud here: https://github.com/carlosmccosta/dynamic_robot_localization/blob/edcbde5ccf9b8cef4256a7651d523a2751e2a8c4/launch/dynamic_robot_localization_system.launch#L133 and the script will predict my localization? Or should I provide some more data?
The reference_pointcloud_filename
is the map that the system loads when it starts.
https://github.com/carlosmccosta/dynamic_robot_localization/blob/noetic-devel/launch/dynamic_robot_localization_system.launch#L32
If your map is not static and changes over time, you can update the reference point cloud by sending a msg to reference_pointcloud_topic
or reference_costmap_topic
https://github.com/carlosmccosta/dynamic_robot_localization/blob/noetic-devel/launch/dynamic_robot_localization_system.launch#L47
https://github.com/carlosmccosta/dynamic_robot_localization/blob/noetic-devel/launch/dynamic_robot_localization_system.launch#L45
The ambient sensor data is received in the ambient_pointcloud_topic
https://github.com/carlosmccosta/dynamic_robot_localization/blob/noetic-devel/launch/dynamic_robot_localization_system.launch#L51
If you have sensors that publish ros sensor_msgs/LaserScan, you can specify the topics in laser_scan_topics
, which will use my laserscan_to_pointcloud
package to assemble and convert the topics to point clouds:
https://github.com/carlosmccosta/dynamic_robot_localization/blob/noetic-devel/launch/dynamic_robot_localization_system.launch#L71
https://github.com/carlosmccosta/laserscan_to_pointcloud/blob/master/launch/laserscan_to_pointcloud_assembler.launch#L4
For every message received in ambient_pointcloud_topic
, the system will try to align it to the reference point cloud and estimate the pose of the robot.
When the system starts, it will try to find the initial pose of the robot using feature matching.
If your environment does not have good features, you can manually specify the initial pose in rviz, which will send a message to /initialpose
https://github.com/carlosmccosta/dynamic_robot_localization/blob/noetic-devel/launch/dynamic_robot_localization_system.launch#L7
http://wiki.ros.org/rviz/UserGuide#A2D_Pose_Estimate_.28Keyboard_shortcut:_p.29
After having a valid initial pose, the system will enter tracking mode using the Iterative Closest Point (ICP) algorithm to align the sensor data with the reference map.
Thank you for the detailed response! I have one more question, how do I visualise the estimated localization in rviz?
Hello,
You can add a pose arrow or axis associated with topic /dynamic_robot_localization/localization_pose
(geometry_msgs/PoseStamped).
http://wiki.ros.org/rviz/DisplayTypes/Pose
Or you can show the map
frame_id in TF.
http://wiki.ros.org/rviz/DisplayTypes/TF
Have a nice day,
Hi, I tried the localization node, but I'm not able to visualise the ambient pointcloud nor the localization pose. What the fixed frame should be? In my ambient_pointcloud_topic
the default fixed frame is rslidar
, but it doesn't work with ambient_pointcloud
. The terminal window shows a message: [pcl::SampleConsensusModelLine::optimizeModelCoefficients] Not enough inliers to refine/optimize the model's coefficients (2)! Returning the same coefficients.
Should I change some parameters, or my pointcloud is just not compatible?
Hello,
The fixed frame should be the one configured in the argument map_frame_id, which by default is map
.
The dynamic_robot_localization_system.launch is configured for 3 DoF (x, y, theta) and expects sensor data from sensor_msgs/LaserScan messages.
If you want to perform 6 DoF localization (x, y, z, roll, pitch, yaw), then you need to configure drl for that use case. In the dynamic_robot_localization_system_6dof.launch you have an example of a possible configuration, which was a refactoring of the tests from ethzasl_kinect_dataset.launch, used to generate this video.
Have a nice day,
I stll have problems with this configuration. I used the dynamic_robot_localization_system_6dof.launch file as you recommended, and changed base_link
to rslidar
. But now the topics don't publish any messages. In my recorded bag file the rslidar
frame actsas a base_link
. But I don't understand how the map
frame works. Is it generated by the launch file?
Hello,
What kind of sensor do you have ? Is it a 2D lidar that outputs sensor_msgs/LaserScan or a 3D lidar that generates sensor_msgs/PointCloud2 ?
If it is a 2d lidar then you are likely performing 3 DoF localization (example here) and you should look at dynamic_robot_localization_system.launch.
If it is a 3d lidar, then you probably want 6 DoF localization (example here) and you should look at dynamic_robot_localization_system_6dof.launch.
The TF package manages matrix transformations between coordinate systems: http://wiki.ros.org/tf2
The REP 105 specifies the naming conventions for mobile robotics: https://www.ros.org/reps/rep-0105.html
Namely:
map -> odom
is published by a global localization node (such as drl) and can have discrete jumps to correct the robot poseodom -> base_link
is published by the wheel / visual / imu odometry node and has smooth evolution over timedrl when used for mobile robotics publishes map -> odom
, more precisely:
map -> odom
TFFor following the conventions, you need to publish the tf that connects odom -> base_link
and base_link -> lidar_frame_id
If you do not have odometry, then you can set the odom_frame_id to an empty string which will cause drl to publish the tf map -> base_link
The default naming is map -> odom -> base_link
, but you can changed the names if you want.
For example, if you do not want to follow the conventions and just want to track you lidar in 3d space, you need to set:
<arg name="map_frame_id" default="map" />
<arg name="odom_frame_id" default="" />
<arg name="base_link_frame_id" default="rslidar" />
<arg name="sensor_frame_id" default="rslidar" />
But keep in mind that you need to specify the reference point cloud file path and the initial pose of the sensor within that reference point cloud, in the dynamic_robot_localization_system_6dof.launch.
You can look at ethzasl_kinect_dataset_high_complexity_slow_fly_movement.launch (and analyze the chain of launch files until reaching the drl node) for an example of 6 DoF tracking of a 3d sensor (video here).
Have a nice day,
Hi, I would like to perform localization of a lidar which is moving in a space, which is also recorded as a pcl pointcloud. I checked the big config file and I haven't seen any option for map upload. Where can I find a good launch file for this? Maybe it would be better to use the
dynamic_robot_localization_tests
package? Regards