LiDAR

LiDAR (Light Detection and Ranging) is an active sensing technology that uses laser light to measure distances and capture the geometry of an environment. A LiDAR scanner supplies its own illumination, allowing it to measure surfaces in darkness. In computer vision, these measurements give machines spatial information for locating objects, navigating environments, and building three-dimensional maps.

How LiDAR Scanners Work#

A pulsed LiDAR sensor emits laser light and measures how long reflected light takes to return. Distance equals the speed of light multiplied by the round-trip travel time, divided by two. Combining each distance with the beam direction produces a surface location. Repeated measurements form a point cloud: a collection of points with x, y, and z coordinates, often accompanied by return intensity, the strength of the reflected signal.

Laser-based LiDAR developed through early optical-radar experiments. Giorgio Fiocco and Louis Smullin demonstrated atmospheric laser measurements in 1963, detecting scattering layers between 60 and 140 kilometers above Earth. Their experiment established an early application of laser echoes to remote measurement.

A 2D scanner samples a plane around a robot; a 3D scanner measures across multiple elevations. Airborne systems additionally combine laser ranges with positioning and orientation measurements to place points in geographic coordinates. Topographic and bathymetric LiDAR serve different environments: topographic instruments typically use near-infrared light for land mapping, while bathymetric instruments use green light to measure underwater terrain.

Turning LiDAR Measurements Into AI Perception#

Point clouds describe geometry. Machine learning adds interpretation through tasks such as 3D object detection, which estimates an object's location, dimensions, and orientation. PointPillars, introduced by Alex H. Lang and colleagues in 2018, groups points into vertical columns, learns features for those columns, and processes the resulting bird's-eye-view representation with a detection network. This converts irregular measurements into a structure suited to efficient convolutional processing.

Sensor fusion combines LiDAR geometry with camera appearance. Images contribute color, texture, and recognizable details; LiDAR contributes measured range and shape. Calibration establishes the rotation and translation between sensors, while synchronization aligns their timestamps. Errors in either step can associate a camera detection with points belonging to another object.

Ultralytics YOLO supports this image-to-geometry workflow through its ROS integration: YOLO26 segmentation masks from aligned RGB images can select corresponding points in an organized, color-bearing point cloud, isolating objects in 3D.

LiDAR, Radar, and Camera-Based Depth#

These technologies provide complementary measurements:

• Depth estimation: Camera-based methods infer distance from images. LiDAR measures range using emitted laser light; a depth map is one possible representation of distance information.
• Radar: Uses radio waves and commonly measures range and radial velocity—the speed toward or away from the sensor. It remains useful in rain and fog that degrade optical sensing.
• 3D reconstruction: Builds a geometric model of a scene. LiDAR supplies measurements that reconstruction systems combine into maps or surfaces.

LiDAR point clouds become sparser with distance, leaving distant pedestrians and other small objects represented by fewer measurements. Fog and snow also scatter laser light, producing backscatter that distorts point clouds and disrupts detection. These effects occur in automotive perception, where missing geometry or weather-generated returns can change which obstacles a detector identifies.

Applications in Driving, Robotics, and Mapping#

• Autonomous driving: The Waymo Driver combines LiDAR, cameras, and radar to perceive vehicles, pedestrians, and cyclists. LiDAR supplies surrounding geometry, while the perception system interprets sensor data to identify road users and their positions.
• Mobile robotics: TurtleBot3's SLAM workflow uses laser scans to build a map while estimating the robot's location. SLAM means simultaneous localization and mapping; repeated observations connect the robot's changing position to persistent environmental features.
• Aerial mapping and environmental monitoring: The National Ecological Observatory Network's airborne LiDAR measures forest structure, including canopy height and vegetation density. A single pulse can produce multiple returns from treetops, branches, understory, and ground because portions of its light pass through gaps in the vegetation.