Field Capture

A point cloud does not come out of the scanner clean. Chapter 7 explained how LiDAR works and what the scanner captures. This chapter covers the field half of the pipeline: the physical work of planning a survey, calibrating the rig, capturing the site, and verifying coverage before the crew leaves. Chapter 9 covers ground control, the survey discipline that decides whether the map agrees with the earth, and Chapter 10 covers what happens at the workstation.

The people who do this work are surveyors, field operators, and processing technicians. On a small team, one person does all three.

Capturing from the robot’s perspective

Every scan must be taken from the same viewpoint the autonomous vehicle will eventually have: at driving height, along the actual travel paths. A scan captured from a drone or a survey tripod records geometry the robot will never see, from angles it will never occupy. For HD map creation, those scans are irrelevant.

The practical question is not which scanner to buy. It is whether a vehicle can reach every surface the robot will eventually drive on.

Vehicle-mounted. The scanner is mounted on a car, cart, or robot at roughly the same height as the autonomous vehicle. This is the default for any surface a vehicle can traverse: roads, open yards, accessible lanes. At Harbor Yard, a vehicle-mounted scanner covered every road lane, the loop road, and the open areas around the warehouse.

Handheld or worn. A portable scanner carried by a person walking the site. Slower than vehicle-mounted, but reaches where a vehicle cannot: narrow corridors, interior spaces, behind stacked obstacles. The operator’s path discipline determines the data quality. Walk steadily, maintain overlap between adjacent passes, and close loops deliberately. At Harbor Yard, a handheld scanner filled in the loading dock interiors and the narrow container canyons where the vehicle could not fit.

The distinction is not the form factor. It is access. A vehicle-mounted scanner and a handheld scanner at the same height produce comparable geometry, with one caveat: a handheld rig rarely carries the high-grade IMU and GNSS integration of a vehicle scanner, so its trajectory drifts faster between fixes. Tight path discipline, deliberate loops, and closely spaced control keep that drift in check. Choose the method that covers every travel surface from the robot’s eye level.

Planning the survey

A mobile scanner never stops moving. There is no tripod and no fixed datum, only a trajectory: the computed path of the vehicle or the person carrying the sensor. The point cloud is only as accurate as the trajectory, which is why the scanner, the IMU, and the GNSS receiver must be tightly synchronized, and why every choice in the plan exists to keep the trajectory honest.

You do not show up at a site with a scanner and figure it out on the day. You plan the survey on a map first. The plan answers three questions.

Where do I put the scanner? The goal is coverage of every surface that will appear in the map: every road lane, every intersection, every dock face, every curb edge. For a vehicle-mounted scanner, this means driving every road segment in both directions, with overlap at the edges so adjacent passes share geometry. Ground-level sensors cast data shadows behind parked vehicles, utility poles, and trash bins; a pass in each direction captures the assets on both sides of the roadway. For a static scanner, it means placing scan positions so that every surface is within range and line-of-sight of at least two positions. One position is not enough: a surface visible from only one angle will have shadows, gaps where the laser could not reach.

Where are my loop closures? A loop closure, returning to a place the scanner has already been, is what lets SLAM recognize the match and wipe out accumulated drift. Chapter 11 covers the mechanics. The survey plan must include loop closures deliberately. At Harbor Yard, the vehicle route starts at the gate, follows the loop road counterclockwise, crosses the bisector, returns along the cross road, and exits through the gate a second time. That final pass through the start point is the loop closure. Without it, the SLAM trajectory drifts and the point cloud is warped. Start and finish in open sky, where satellite visibility is good, and let the system sit still for a few minutes at the start so it can initialize before it moves. Steady driving matters too: sudden acceleration, hard braking, and swerving degrade the IMU and stretch or compress the cloud. A consistent, moderate speed produces cleaner data than a hurried one.

Where are my ground control points or fixed landmarks? Fixed landmarks matter for every map: they are the distinctive geometry a robot will grab onto when it localizes against the finished cloud, which Chapter 12 explains in full. In the field they are tie points: features already in the scene, street lights, stop signs, a corner of the dock, the gate, that carry the anchor into the point cloud. A tie point must be visible from at least one scan position, ideally two or more, and each one gets its world coordinates from the scanning rig’s own RTK receiver during the drive. Ground control itself does not have to be visible to the scanner at all. Whether the application needs an earth anchor is a question the next chapter works through; when it does, a survey nail driven into the earth provides it, and the nail’s job is simply to give the RTK base station known coordinates. The rig’s RTK-corrected trajectory inherits that anchor and ties the cloud to it. Where the scene offers no usable landmarks, paint high-contrast targets on the pavement, chevrons or L shapes, and survey them with an RTK rover or a total station. Spacing follows GNSS quality: every 100 to 150 m in open sky, every 50 m where buildings or container stacks block satellites, and every 25 to 50 m for handheld work under canopy or indoors, where the IMU runs blind. If a chosen landmark is blocked by a parked truck or a stacked container on survey day, you need a backup.

Calibrating the rig

Before project data is collected, the rig itself has to be calibrated.

Boresighting. The exact offsets between the LiDAR mirror, the IMU center, and the GNSS antenna, the lever arm, plus the small angular misalignments between their frames, are solved once, before capture, not per survey. The standard routine is a figure-eight drive over a feature-rich area, a parking lot with distinct paint stripes and building corners, where the overlapping scans give the software enough constraint to compute the geometry. A rig that has not been boresighted produces clouds that are subtly warped in ways registration cannot fix later.

Time synchronization. Every sensor must be stamped to a common clock, usually GNSS time distributed by a pulse-per-second signal. Loose synchronization smears geometry, and it shows up first in colorization: a rolling-shutter camera beside the LiDAR exposes each frame line by line, so the software must apply per-line time offsets to compensate for vehicle motion. Without that correction, the color draped over the point cloud lands a few centimeters from the geometry it belongs to.

What goes wrong on site

The scan plan assumes a static world. The real site has weather, vehicles, and people.

Moving objects. A forklift crossing the yard during a scan becomes a streak: a trail of points where the laser caught it at successive positions. A person walking becomes a smeared ghost. Birds in flight produce isolated clusters hanging in mid-air. Most of these are removable with automated filters, but dense traffic can obscure the ground surface entirely. If a truck was parked in front of a dock face for the entire scan, that dock face is missing from the point cloud. The only fix is to wait for the truck to move, or to return later.

Weather. Rain scatters LiDAR pulses. A light drizzle produces noisy returns; heavy rain produces a fog of false points within a few meters of the scanner. Fog and dust do the same. Most field crews will not scan in rain. Direct sunlight is less obvious but still problematic: it heats surfaces unevenly, creating thermal distortion that affects long-range accuracy, and it blinds cameras that are used alongside the LiDAR for colorization. Overcast days produce the cleanest data.

Reflections. Glass windows, polished metal, and standing water reflect the laser pulse away from the scanner. The return never arrives, and the surface is invisible in the point cloud. Worse, the reflected pulse may strike a real surface and return from an impossible angle, creating a phantom point behind the reflective surface. A warehouse with large windows can appear to have interior walls floating 20 m behind the actual building. Recognizing these as artifacts requires experience.

Verifying coverage before you leave

Before the crew packs up, someone checks that the data is complete. This does not require processing the full point cloud. Most capture systems include a real-time preview: a low-resolution visualization of accumulated points overlaid on the site plan.

The operator checks four things:

1. Every road surface is covered. No gaps longer than a meter in any lane the robot will drive.
2. Every planned landmark is visible, and the trajectory held RTK fix where it matters. Each chosen landmark appears in at least one scan, and the trajectory log shows fix wherever the scanner passed a planned landmark. If a landmark is missing, the operator checks whether it was physically blocked and, if possible, clears the obstruction and re-scans the area.
3. Loop closures were actually driven. The trajectory preview confirms the route returned through the start point.
4. No obvious artifacts dominate any region. A region that is solid noise instead of geometry, a swath of phantom points, a section where the sensor reported errors instead of returns.

If any check fails, the crew re-scans the affected area while still on site. Ten minutes of re-scanning saves a full day of return visit.

What comes next

The next chapter covers ground control: how the coordinates of ground control points get measured, why the answer is usually RTK, and whether the application needs an earth anchor at all.