# Pointcloud Processing > --- name: pointcloud-processing description: 'Process PointCloud2 data in ROS 2. Use when filtering, downsampling, or segmenting 3D point clouds with PCL.' --- # PointCloud2 Processing ## Binary Data Format `sensor_msgs/msg/PointCloud2` stores 3D point data in a tightly packed binary buffer. Key fields: | Field | Type | Description | |---|---|---| | `header` | Header | Frame and timestamp | | `height` | uint32 | 1 for unorganized clouds; rows for organized (e.g., depth camera) | | `width` | uint32 | Number of points per row | | `fields[]` | PointField[] | Description of each channel (name, offset, datatype, count) | | `is_bigendian` | bool | Byte order of the data | | `point_step` | uint32 | Bytes per point | | `row_step` | uint32 | Bytes per row (`point_step * width`) | | `data` | uint8[] | The raw binary point data | | `is_dense` | bool | `true` if no invalid points (no NaN/inf) | Common field layouts: - **XYZ**: fields `x`, `y`, `z` — each `FLOAT32`, `point_step` = 12 or 16 (with padding) - **XYZRGB**: adds `rgb` field packed as a `FLOAT32` (actually 3 bytes: R, G, B) - **XYZI**: adds `intensity` as `FLOAT32` ## Reading PointCloud2 in Python ```python import sensor_msgs_py.point_cloud2 as pc2 def cloud_callback(msg): for point in pc2.read_points(msg, field_names=('x', 'y', 'z'), skip_nans=True): x, y, z = point ``` In C++, use `sensor_msgs::PointCloud2Iterator` or the `pcl_conversions` bridge to convert to `pcl::PointCloud`. ## PCL-ROS Filtering Pipeline The `pcl_ros` package wraps PCL filters as ROS 2 nodelets/components. Key filters: ### VoxelGrid Downsampling Divides space into voxels of size `leaf_size` and replaces all points in each voxel with their centroid. Reduces point count dramatically while preserving structure. ```yaml voxel_grid: ros__parameters: leaf_size_x: 0.05 leaf_size_y: 0.05 leaf_size_z: 0.05 input_topic: /camera/points output_topic: /camera/points_downsampled ``` ### PassThrough Filter Crops points to an axis-aligned bounding box. Remove floor, ceiling, or limit range: ```yaml passthrough_z: ros__parameters: filter_field_name: "z" filter_limit_min: 0.1 # above ground plane filter_limit_max: 1.5 # below ceiling filter_limit_negative: false ``` ### StatisticalOutlierRemoval For each point, computes the mean distance to its K nearest neighbors. Points with mean distance beyond `stddev_mul * global_stddev` are removed. Effective against sparse noise. ```yaml statistical_outlier: ros__parameters: mean_k: 50 stddev_mul_thresh: 1.0 ``` ### CropBox Similar to PassThrough but operates on a full 3D box with optional transform frame. Useful for removing the robot body from the point cloud: ```yaml crop_box: ros__parameters: min_x: -0.3 min_y: -0.3 min_z: -0.1 max_x: 0.3 max_y: 0.3 max_z: 0.5 negative: true # remove points INSIDE the box input_frame: base_link ``` ## pointcloud_to_laserscan Projects a 3D point cloud into a virtual 2D `LaserScan` for consumption by Nav2 costmaps. Each angular bin takes the nearest (or smallest Z within height range) point. ### Configuration ```yaml pointcloud_to_laserscan_node: ros__parameters: target_frame: base_link # transform cloud to this frame first transform_tolerance: 0.01 min_height: 0.05 # ignore points below (ground) max_height: 1.0 # ignore points above (ceiling/irrelevant) angle_min: -3.14159 # full 360° or restrict to FOV angle_max: 3.14159 angle_increment: 0.00436 # ~0.25° resolution scan_time: 0.1 range_min: 0.15 range_max: 10.0 use_inf: true # emit inf for empty bins inf_epsilon: 1.0 ``` ### Launch Example ```python Node( package='pointcloud_to_laserscan', executable='pointcloud_to_laserscan_node', name='pc_to_scan', remappings=[ ('cloud_in', '/camera/points_filtered'), ('scan', '/camera/scan'), ], parameters=[pointcloud_to_laserscan_params], ), ``` ## Frame Considerations Point clouds from depth cameras arrive in the camera's optical frame (Z-forward, X-right, Y-down). The `pointcloud_to_laserscan` node uses TF to transform to `target_frame` (usually `base_link`) before projection. Ensure TF is published between `camera_depth_optical_frame` → `base_link` via your URDF static transforms. ## Integration with Nav2 After projection to a virtual scan, add it as an observation source in the costmap: ```yaml observation_sources: lidar_scan depth_scan depth_scan: topic: /camera/scan data_type: LaserScan marking: true clearing: true obstacle_max_range: 5.0 obstacle_min_range: 0.15 ``` For direct point cloud sources (without `pointcloud_to_laserscan`), use `data_type: PointCloud2` in the costmap observation source and configure `min_obstacle_height`/`max_obstacle_height`.