How lidar mapping works

01 THE EQUATION THAT MAKES LIDAR POSSIBLE

Lidar determines the distance to a surface with a formula that fits on one line: d equals c times t divided by two, where d is the distance, c is the speed of light, and t is the time for a laser pulse to travel to the target and back. The division by two accounts for the round trip. The formula is simple. The difficulty is in measuring t.

Light travels at roughly 300,000 kilometers per second, which means it covers 30 centimeters in one nanosecond. A round trip to an object 15 meters away takes 100 nanoseconds. To resolve distance to 1.5 centimeters, the system must measure time to 100 picoseconds, one tenth of a nanosecond. That timing precision is the engineering core of every lidar system, and it is why lidar hardware costs what it costs.

Two detection schemes exist. Direct time-of-flight lidar measures the round-trip time directly by sending a short pulse and timestamping the return. Coherent lidar, used in frequency-modulated continuous wave systems, sweeps the laser's frequency over time and measures the beat frequency between the transmitted and received signals. The beat frequency encodes the distance. Coherent detection is more sensitive and more robust to ambient light but requires more complex optics and electronics.

02 THE FOUR COMPONENTS THAT MAKE A LIDAR

A lidar system is built from four subsystems, and each constrains the others. The laser generates the light pulse. The scanner and optics direct that pulse across the scene. The photodetector and receiver electronics capture the returning light and convert it to an electrical signal. The position and navigation system records where the sensor was pointing and where the sensor was located when each measurement was taken.

The laser's wavelength is a first design choice. Most non-scientific lidars use 600 to 1,000 nanometer lasers because detectors for that range are mature and inexpensive. 1,550 nanometer lasers are eye-safe at higher power levels, since this wavelength is strongly absorbed by water and barely reaches the retina, allowing longer-range operation without blinding risk. Airborne topographic lidars typically use 1,064 nanometer YAG lasers, while bathymetric systems that map underwater terrain use 532 nanometer frequency-doubled lasers because 532 nanometers penetrates water with less attenuation.

The scanner determines how the laser beam is steered across the scene. Mechanical scanners use rotating mirrors or polygons to sweep the beam. Solid-state scanners use micro-electromechanical mirrors or optical phased arrays that steer the beam without moving parts. The scanner's speed determines the frame rate: how quickly the lidar can build a complete scan of its environment. A rotating mirror that completes 10 revolutions per second captures 10 frames per second, and each frame contains as many points as the laser can fire within one revolution.

The photodetector must be fast enough to resolve the return pulse and sensitive enough to detect the small fraction of light that returns from a distant surface. Avalanche photodiodes, which multiply the incoming photon signal through an internal gain stage, are common in high-performance systems. The detector's bandwidth determines the minimum pulse width the system can resolve, which in turn determines the range resolution: better bandwidth means the system can distinguish two surfaces that are close together in depth.

03 FROM DISTANCE MEASUREMENTS TO POINT CLOUD

A single lidar measurement produces one distance along one direction. To build a map, the system must take many measurements from many directions, and each measurement must be tagged with the angle it was taken at. A scanning lidar sweeps its laser beam across the scene, firing thousands to millions of pulses per second. Each pulse produces one return, and each return is converted to a 3D point using the measured distance and the known beam angle.

The result is a point cloud: a set of points in three-dimensional space, each with Cartesian coordinates x, y, and z, and often with additional attributes like reflectance intensity, the amplitude of the return signal, or color from a co-located camera. A point cloud is not a surface. It is a set of samples from surfaces, and the density of those samples determines how much of the scene the cloud represents.

The density depends on three factors: the laser's pulse repetition rate, the scanner's angular resolution, and the distance to the target. A lidar that fires 200,000 pulses per second with a 0.1-degree angular increment will produce roughly 360,000 points per 360-degree scan line. At 10 scan lines per second from a rotating mirror, that is 3.6 million points per second. But the points spread as they travel outward: a 0.1-degree beam at 100 meters is separated by about 17 centimeters from its neighbor, which means small objects at long range may be missed entirely.

04 SLAM AND THE MAP THAT BUILDS ITSELF

A stationary lidar produces a static point cloud of what it can see from where it is. A moving lidar, mounted on a vehicle, a drone, or a robot, must produce a continuous map of an environment it is traveling through. The problem is that the sensor does not know where it is, and it does not know the shape of the environment. It must solve both simultaneously. This is the problem of simultaneous localization and mapping, or SLAM.

SLAM appears to be a chicken-and-egg problem: to map the environment, the system must know its position; to know its position, the system must have a map. The solution is to build both incrementally. The system takes a scan, moves, takes another scan, and finds the overlap between the two scans. The overlap tells the system how far it moved and how the environment is shaped. Each new scan refines the map and the position estimate simultaneously.

The algorithms that solve SLAM are drawn from computational geometry and probabilistic estimation. Extended Kalman filters represent the map and the robot's pose as a probability distribution, updating both as new measurements arrive. Particle filters maintain a set of hypotheses about the robot's position, weighting each by how well it explains the observations. Graph-based SLAM constructs a graph of pose nodes and measurement constraints, then optimizes the entire graph to find the most consistent trajectory and map. Each approach trades computational cost against accuracy and robustness, and the choice depends on the environment and the available computing power.

05 AIRBORNE LIDAR AND MAPPING FROM ABOVE

Airborne lidar is the tool that made large-scale topographic mapping practical. A lidar mounted on an aircraft or satellite fires pulses downward, and the returning signals map the terrain below. An aircraft can map terrain at 30-centimeter resolution or better, far exceeding what photogrammetry alone can achieve in vegetated areas, because laser pulses penetrate gaps in the canopy and return from the ground beneath.

The key innovation that made airborne lidar useful for mapping is the GPS and inertial measurement unit. An aircraft in flight is constantly moving and rotating, and every lidar measurement must be corrected for the aircraft's position and orientation at the instant the pulse was fired. The GPS provides the position, and the inertial measurement unit provides the orientation. Together, they transform each raw distance measurement, taken from a moving platform, into a georeferenced point in a fixed coordinate system.

Airborne lidar has transformed archaeology as much as geography. Dense forests that hid ancient cities from aerial photography are transparent to lidar, which penetrates the canopy and maps the ground surface beneath. Lidar surveys have revealed tens of thousands of structures at sites like Angkor in Cambodia and Nan Madol in Micronesia, reshaping the understanding of pre-industrial urban extent. The technology that maps a road also maps a lost city.

06 LIDAR IN AUTONOMOUS VEHICLES

The most familiar lidar application is the autonomous vehicle. A self-driving car uses lidar to detect obstacles, map lane markings, and build a real-time 3D representation of the driving environment. The lidar's point cloud output provides the necessary data for the vehicle's software to determine where obstacles exist and where the vehicle is in relation to those obstacles.

Lidar was the key enabler behind Stanley, the vehicle that won the 2005 DARPA Grand Challenge, the first autonomous vehicle to successfully complete the course. Five Velodyne lidar units on the roof of the winning vehicle produced a 360-degree view of the environment that no camera-based system could match at the time. The lidar did not just see obstacles. It measured their distance and shape, and that measurement was what made autonomous navigation possible.

Current automotive lidars use rotating hexagonal mirrors that split the laser beam into multiple channels. Upper beams detect vehicles and obstacles ahead, while lower beams detect lane markings and road features. The point cloud is fed into a perception pipeline that segments objects, classifies them, and tracks their motion over time. The lidar does not identify what an object is. It tells the vehicle where the object is and what shape it is, and the perception system does the rest.

07 THE LIMITS OF LIDAR MAPPING

Lidar is powerful, but it is not omniscient. The laser pulse reflects from the first surface it hits, which means lidar cannot see through solid objects. A pedestrian behind a truck is invisible to a lidar on a car, because the pulse returns from the truck, not the pedestrian. Rain, fog, and snow scatter the laser beam, reducing range and adding noise to the point cloud. Black surfaces absorb most of the light, producing weak returns that may fall below the detection threshold. Mirrors and glass redirect the beam, producing phantom returns at incorrect distances.

The point cloud itself is not a semantic map. A lidar returns points, but it does not label them. The point that represents a pedestrian and the point that represents a fire hydrant look the same to the sensor until a classification algorithm processes the cloud. The classification, which uses the geometry of local point clusters, is a machine learning problem layered on top of the physics of time-of-flight. The lidar provides the raw 3D data; the perception system provides the meaning.

The cost trajectory of lidar has been steep. In 2007, a Velodyne lidar for autonomous driving cost roughly 80,000 dollars. In 2026, solid-state lidars for automotive applications cost hundreds of dollars, and some consumer devices, like the iPhone, include lidar sensors that map rooms for augmented reality applications. The price drop has moved lidar from a specialty tool for survey aircraft and robot cars to a sensor that fits in a pocket, and the mapping applications that follow from that ubiquity are still being discovered.