Understanding LiDAR Data: Point Clouds, Formats and ROS
Point clouds, intensity, multiple echoes, PCD/LAS/ROSbag formats — everything you need to know to work with raw LiDAR sensor data.
What Is a Point Cloud?
A point cloud is the native output format of a LiDAR sensor. Each point represents a laser shot that hit a surface and returned to the detector. A typical point cloud contains 50,000 to 2.5 million points per second depending on the sensor.
Each point carries several attributes: position (x, y, z), return intensity, sometimes arrival time and channel number. Some sensors add RGB color, classification or echo number.
Intensity and Reflectivity
Intensity measures the strength of the returned signal. A white surface reflects more than a black surface. Intensity helps distinguish materials but is not absolutely calibrated: it depends on distance, incidence angle and sensor settings.
Multiple Echoes (Multi-Return)
A laser shot can partially pass through foliage and hit a branch then the ground: two returns are recorded (first and last return). Multi-echo LiDARs (RIEGL, some Ouster) record up to 5 returns per shot — essential for forestry mapping.
Common File Formats
PCD (PCL): open format for robotics and perception. LAS/LAZ (ASPRS): compressed geospatial standard with classification. PLY: supports color and normals for computer vision. ROSbag: ROS container for recording and replay.
LiDAR Data in ROS
ROS uses sensor_msgs/PointCloud2 for point clouds. Self-describing format with fields x, y, z, intensity, ring, timestamp. Some sensors also publish /scan (LaserScan) for 2D and /imu_raw for inertial data. Temporal sync is critical for SLAM.
Common Noise and Artifacts
Ghost points on reflective surfaces, solar noise in bright sunlight, rain/fog creating parasitic points, and edge effects. Most SDKs include filters: outlier removal, intensity filter, distance threshold and FoV mask.
From Data to Information
Segmentation (grouping points of the same object), classification (labeling: ground, building, vehicle), tracking (temporal association) and SLAM (correlating clouds to build a map). Format choice depends on the application: LAS for mapping, PointCloud2/ROS for robotics, PLY/PCD for AI.