LiDAR and Spatial Maps

A LiDAR scanner does not take pictures. It fires laser pulses and measures how long they take to return. Each return becomes one coordinate in 3D space, with an intensity value that tells you how reflective the surface was. After a scan, you do not have an image. You have millions of distance measurements that record, with centimeter precision, where the physical surfaces are.

LiDAR is the foundation of every HD map. Cameras guess at depth from pixels. Radar gives you blobs. Only LiDAR tells you exactly where everything is, in coordinates you can build on.

How LiDAR works

LiDAR stands for Light Detection and Ranging. The principle is time-of-flight: send a laser pulse, measure how long it takes to return, compute the distance. Because the speed of light is a constant you can generally trust, the round-trip time tells you exactly how far away the reflecting surface is.

Imagine standing at the edge of Harbor Yard. You fire a laser pulse at a stack of shipping containers near the warehouse. Your sensor records the nanosecond the pulse leaves and the nanosecond the reflection returns. The time difference, multiplied by the speed of light and divided by two, gives you the distance. Repeat this 100,000 to 2,000,000 times per second as the sensor spins, and you have a 3D measurement of the yard.

The engineering that makes this work at highway speeds, from a moving platform, with centimeter accuracy is not simple. But from the map builder’s perspective, you need to understand three numbers on the spec sheet.

Density. How many points per square meter. A dense point cloud, 1,000 points per square meter or more, lets you see fine details like curb edges, lane markings, and small obstacles. A sparse point cloud, 10 points per square meter, is fine for building outlines but not for lane-level mapping.

Accuracy. How close each point is to the true surface position. Survey-grade scanners achieve 2 to 5 mm accuracy. Mobile mapping scanners, the kind used to capture Harbor Yard, are typically accurate to 2 to 5 cm. Consumer-grade sensors, like the LiDAR in some phones, can be off by 10 cm or more¹.

Range. How far the laser can reach. Short-range scanners (up to 100 m) are typical for ground vehicles. Long-range scanners (200 m and beyond) are used for airborne and large-site capture. The range affects both density and accuracy, because the laser spot spreads out as it travels.

What the scanner captures

A LiDAR scanner records geometry and nothing else. Each point marks where a laser beam struck something: the scanner fires a pulse, the pulse hits the first solid object in its path and returns, and the scanner records the distance, the angle, and the intensity. After processing, that becomes an (X, Y, Z) coordinate in the scanner’s frame.

What the scanner does not capture is structure. A point cloud has no concept of a surface. To the scanner, the world is a cloud of fireflies: isolated points hanging in space where a beam hit something. The understanding that certain groups of points form an impenetrable road or a solid wall is not measured. It is inferred, by algorithms and humans who recognize that the empty space between those points is occupied, not void.

Meaning is another layer removed. A cluster of points at ground level could be asphalt, concrete, gravel, or painted metal. The scanner records only geometry and reflectivity. Interpreting that geometry as “road,” “sidewalk,” or “loading dock” is the job of semantic labeling, several stages later in the pipeline.

Two kinds of sensors

Sensors on an autonomous vehicle fall into two categories.

Proprioceptive sensors measure the internal state of the robot. Inertial measurement units (IMUs) track acceleration and angular velocity. Wheel encoders count rotations. These sensors tell the robot what it is doing, but nothing about the world outside.

Exteroceptive sensors measure the external environment. LiDAR, cameras, and radar are all exteroceptive. They tell the robot what is around it.

HD mapping depends on exteroceptive sensors, but proprioceptive sensors are essential for making sense of the exteroceptive data. An IMU fills the gaps between LiDAR scans, tracking the scanner’s motion so that points captured at different moments can be placed in a common reference frame. Without that motion compensation, a scan taken from a moving vehicle is warped beyond use. Chapter 11 explains how this works in detail.

Why GNSS is not enough

You might wonder why HD mapping exists at all. Why not put a GNSS receiver on the robot and call it done?

Consumer GNSS is accurate to within a few meters in open sky. In the environments where autonomous vehicles operate, it is often much worse. Urban canyons between tall buildings bounce signals off glass and steel. Metal shipping containers, like the ones stacked at Harbor Yard, reflect GNSS signals in every direction. A robot relying on GNSS alone in the container zone could be off by 10 m, or 20 m, and have no way to know it.

Differential GNSS, which corrects errors against a fixed base station, can reach centimeter accuracy. But it requires a clear view of the sky and a reliable radio link to the base station. In a warehouse, a tunnel, or a dense urban core, it still fails.

HD maps exist because GNSS is not reliable enough to answer the question “where am I, right now, to within 5 cm?” LiDAR can answer that question, but only if it has a map to compare itself against. Building that map is what this book is about.

Direct and indirect methods

Processing philosophies for sensor data split into two camps.

Indirect methods extract a small set of distinctive features or keypoints from the raw data and work with those. They are fast and compact, ideal for limited compute platforms. Think of an indirect method as a scout who returns from Harbor Yard with notes on where the big cranes are: enough to navigate, not enough to reconstruct the site.

Direct methods process all available data, every point, every intensity value. They are potentially more accurate but computationally expensive. The direct method scout tries to photograph every pebble.

LiDAR mapping heavily favors direct methods because the data is already geometric. You are measuring distances directly, not inferring them from pixels. But processing millions of points in real time requires serious compute. Modern SLAM (simultaneous localization and mapping) systems offload the heavy lifting to GPUs, which excel at the parallel number-crunching that point cloud processing demands.

Point cloud formats

The point cloud is the container that holds all those distance measurements. The most common formats:

LAS / LAZ. The industry standard for LiDAR data. LAS is the binary format; LAZ is the compressed version. Nearly every scanner and processing tool reads and writes LAS.

PCD. The Point Cloud Data format, common in the ROS ecosystem. Simpler than LAS, less metadata, easier to parse.

E57. A vendor-neutral format for 3D imaging data, common in surveying and construction. Supports both point clouds and images in a single file.

For HD mapping work, you will most often work with LAS files from the scanner, convert to PCD for ROS-based processing, and archive in LAZ to save space.

What comes next

The next chapter covers field capture: how to plan a survey and capture the scans. Chapters 9 and 10 cover ground control and scan processing, and Chapter 11 then explains SLAM, the computational engine that turns those raw scans into a registered point cloud.

Footnotes

1. The accuracy numbers in this chapter are typical for 2026. Like everything in sensing, they improve every year. The principle does not change: know your scanner’s spec sheet before you trust its output. ↩