lidarslam_ros2

ros2 slam package of the frontend using OpenMP-boosted gicp/ndt scan matching and the backend using graph-based slam.

mobile robot mapping

Green: path with loopclosure
(the 25x25 grids in size of 10m × 10m)

Red and yellow: map

summary

lidarslam_ros2 is a ROS2 package of the frontend using OpenMP-boosted gicp/ndt scan matching and the backend using graph-based slam.
I found that even a four-core laptop with 16GB of memory could work in outdoor environments for several kilometers with only 16 line LiDAR.
(WIP)

requirement to build

You need ndt_omp_ros2 for scan-matcher

clone (If you forget to add the --recursive option when you do a git clone, run git submodule update --init --recursive in the lidarslam_ros2 directory)

cd ~/ros2_ws/src
git clone --recursive https://github.com/rsasaki0109/lidarslam_ros2

build

cd ~/ros2_ws
colcon build --cmake-args -DCMAKE_BUILD_TYPE=Release

io

frontend(scan-matcher)

• input
  /input_cloud (sensor_msgs/PointCloud2)
  /tf(from "base_link" to LiDAR's frame)
  /initial_pose (geometry_msgs/PoseStamed)(optional)
  /imu (sensor_msgs/Imu)(optional)
  /tf(from "odom" to "base_link")(Odometry)(optional)

• output
  /current_pose (geometry_msgs/PoseStamped)
  /map (sensor_msgs/PointCloud2)
  /path (nav_msgs/Path)
  /tf(from "map" to "base_link")
  /map_array(lidarslam_msgs/MapArray)

backend(graph-based-slam)

• input
  /map_array(lidarslam_msgs/MapArray)

• output
  /modified_path (nav_msgs/Path)
  /modified_map (sensor_msgs/PointCloud2)

• srv
  /map_save (std_srvs/Empty)

pose_graph.g2o and map.pcd are saved in loop closing or using the following service call.

ros2 service call /map_save std_srvs/Empty

params

• frontend(scan-matcher)

Name	Type	Default value	Description
registration_method	string	"NDT"	"NDT" or "GICP"
ndt_resolution	double	5.0	resolution size of voxel[m]
ndt_num_threads	int	0	threads using ndt(if 0 is set, maximum alloawble threads are used.)(The higher the number, the better, but reduce it if the CPU processing is too large to estimate its own position.)
gicp_corr_dist_threshold	double	5.0	the distance threshold between the two corresponding points of the source and target[m]
trans_for_mapupdate	double	1.5	moving distance of map update[m]
vg_size_for_input	double	0.2	down sample size of input cloud[m]
vg_size_for_map	double	0.05	down sample size of map cloud[m]
use_min_max_filter	bool	false	whether or not to use minmax filter
scan_max_range	double	100.0	max range of input cloud[m]
scan_min_range	double	1.0	min range of input cloud[m]
scan_period	double	0.1	scan period of input cloud[sec](If you want to compound imu, you need to change this parameter.)
map_publish_period	double	15.0	period of map publish[sec]
num_targeted_cloud	int	10	number of targeted cloud in registration(The higher this number, the less distortion.)
debug_flag	bool	false	Whether or not to display the registration information
set_initial_pose	bool	false	whether or not to set the default pose value in the param file
initial_pose_x	double	0.0	x-coordinate of the initial pose value[m]
initial_pose_y	double	0.0	y-coordinate of the initial pose value[m]
initial_pose_z	double	0.0	z-coordinate of the initial pose value[m]
initial_pose_qx	double	0.0	Quaternion x of the initial pose value
initial_pose_qy	double	0.0	Quaternion y of the initial pose value
initial_pose_qz	double	0.0	Quaternion z of the initial pose value
initial_pose_qw	double	1.0	Quaternion w of the initial pose value
use_odom	bool	false	whether odom is used or not for initial attitude in point cloud registration
use_imu	bool	false	whether 9-axis imu(Angular velocity, acceleration and orientation must be included.) is used or not for point cloud distortion correction.(Note that you must also set the scan_period.)
debug_flag	bool	false	Whether or not to display the registration information

• backend(graph-based-slam)

Name	Type	Default value	Description
registration_method	string	"NDT"	"NDT" or "GICP"
ndt_resolution	double	5.0	resolution size of voxel[m]
ndt_num_threads	int	0	threads using ndt(if 0 is set, maximum alloawble threads are used.)
voxel_leaf_size	double	0.2	down sample size of input cloud[m]
loop_detection_period	int	1000	period of searching loop detection[ms]
threshold_loop_closure_score	double	1.0	fitness score of ndt for loop clousure
distance_loop_closure	double	20.0	distance far from revisit candidates for loop clousure[m]
range_of_searching_loop_closure	double	20.0	search radius for candidate points from the present for loop closure[m]
search_submap_num	int	2	the number of submap points before and after the revisit point used for registration
num_adjacent_pose_cnstraints	int	5	the number of constraints between successive nodes in a pose graph over time
use_save_map_in_loop	bool	true	Whether to save the map when loop close(If the map saving process in loop close is too heavy and the self-position estimation fails, set this to false.)

demo

trial environment

demo data(ROS1) is hdl_400.bag in hdl_graph_slam
The Velodyne VLP-32 was used in this data.
To use rosbag in ROS1, use rosbag2_bag_v2

rviz2 -d src/lidarslam_ros2/lidarslam/rviz/mapping.rviz 

ros2 launch lidarslam lidarslam.launch.py

ros2 bag play -s rosbag_v2 hdl_400.bag 

Green: path with loopclosure, Yellow: path without loopclosure

The larger environment

demo data(ROS1) by Autoware Foundation
https://data.tier4.jp/rosbag_details/?id=212
The Velodyne VLP-16 was used in this data.

rviz2 -d src/lidarslam_ros2/lidarslam/rviz/mapping_tukuba.rviz 

ros2 launch lidarslam lidarslam_tukuba.launch.py

ros2 bag play -s rosbag_v2 tc_2017-10-15-15-34-02_free_download.bag 

Green: path

Red and yellow: map

Used Libraries

• Eigen
• PCL(BSD3)
• g2o(BSD2 except a part)
• ndt_omp (BSD2)

Related packages

• li_slam_ros2 (BSD2)