ROS 2 LiDAR obstacle detection

Stack of two packages in this repository:

Package	Role
lidar_obstacle_detection	Node lidar_cloud_ingress: subscribe PointCloud2, rigid transform into the robot base frame, spatial filter + voxel, optional temporal merge, optional surface segmentation + DBSCAN → ObstacleList (with YZ-smoothed closest_surface_point) and debug topics (optional depth-shaded segmented cloud, optional closest-point RViz spheres).
lidar_obstacle_detection_msgs	Message definitions (Obstacle, ObstacleList).

Defaults in lidar_obstacle_detection/config/lidar_cloud_ingress.yaml target a Unitree Go2 setup: input cloud /utlidar/cloud, output in base_link, and mount extrinsics matching that robot. The published cloud is always expressed in output_cloud_frame_id (usually base_link). A static TF base_link → lidar_link is published only for visualization alignment in RViz; it does not change the point math.

---

Figures

RViz2 — point cloud in base_link, segmentation display, obstacle markers (bounding box + normal), and TF frames (base_link, lidar_link):

Pipeline — raw scan → intermediate processing → clustered obstacles (cosine similarity surface split + DBSCAN):

---

Build

From your ROS 2 workspace (parent of this src tree):

cd ~/ros2_ws
rosdep install --from-paths src/object_detection --ignore-src -r -y   # optional; first-time deps
python3 -m pip install -r src/object_detection/lidar_obstacle_detection/requirements.txt
colcon build --packages-up-to lidar_obstacle_detection --symlink-install
source install/setup.bash

---

Launch

ros2 launch lidar_obstacle_detection lidar_cloud_ingress.launch.py

• With RViz (default): omit headless or set headless:=false.
• Headless (no RViz, e.g. on-robot or SSH): headless:=true.

Override parameters file: ingress_params_file:=/path/to/your.yaml. Verbose logs: verbose:=true.

---

Customizing for another robot / LiDAR

1. Config YAML (required)

Edit lidar_obstacle_detection/config/lidar_cloud_ingress.yaml (or a copy passed as ingress_params_file). Fields that almost always change for a new platform:

Parameter	What to set
input_topic	Your driver’s sensor_msgs/PointCloud2 topic.
output_cloud_frame_id	Frame of the processed cloud (typically the robot base / base_link).
lidar_mount_tf_parent_frame	Must match output_cloud_frame_id so TF and cloud headers agree.
lidar_link_frame	Child name for the visualization static TF (often matches URDF LiDAR link).
lidar_mount_tf_xyz, lidar_mount_tf_rpy_rad	Parent → LiDAR extrinsic (meters, radians); same geometry the node uses to transform points.
driver_cloud_to_lidar_link_xyz, driver_cloud_to_lidar_link_rpy_rad	Extra rigid step if the driver’s points are not already in the frame your URDF expects before the mount (often all zeros).
publish_lidar_mount_static_tf	Set false if another node (e.g. robot_state_publisher) already publishes the same transform.

Then tune fov_deg, max_depth, voxel_size, temporal_*, and perception (dbscan_*, cosine_threshold, etc.) for your sensor density and environment. For forward depth in base_link (+X ahead), yz_depth_* parameters control YZ-bin smoothing of the minimum-X surface estimate; perception_segmented_cloud_depth_shading modulates segmented obstacle colors by that smoothed depth in RViz; publish_closest_surface_markers adds sphere markers at the chosen closest points (see the package README).

2. RViz (optional but typical)

Saved config: lidar_obstacle_detection/rviz/lidar_ingress.rviz.

• Set Global Options → Fixed Frame to your base frame (default base_link).
• Point each display’s Topic at the names you set in YAML (output_topic, colored_segmented_cloud_topic, obstacle_markers_topic, etc., if you renamed them). With perception_segmented_cloud_depth_shading, the segmented cloud encodes smoothed forward depth in brightness; enable publish_closest_surface_markers to add closest-surface spheres on the obstacle MarkerArray topic (marker namespace lidar_obstacle_closest_surface).

---

Further detail

See the package README: lidar_obstacle_detection/README.md (topics, QoS, algorithm layout, message README link).

---

Citation

Related work was poster-presented at ICRA 2025, Workshop of Field Robotics. If you use this software or ideas from that line of work, please cite:

T. Girgin, E. Girgin, and C. Kilic, “Learning Rock Pushability on Rough Planetary Terrain,” arXiv preprint arXiv:2505.09833, 2025. https://arxiv.org/abs/2505.09833

BibTeX:

@article{girgin2025learning,
  title={Learning Rock Pushability on Rough Planetary Terrain},
  author={Girgin, Tuba and Girgin, Emre and Kilic, Cagri},
  journal={arXiv preprint arXiv:2505.09833},
  year={2025}
}