How 3D Point Cloud Processing Works

TL;DR

3D point cloud processing turns the billions of raw points captured by a laser scanner or LiDAR drone into clean, georeferenced, measurable data — through registration, noise removal, classification, georeferencing to GDA2020/MGA2020, and finally modelling into CAD or BIM deliverables. The field capture is only the first hour of the job; processing is where a usable 2–6 mm dataset is actually produced, and it is where most of the value (and most of the errors) live.

Key takeaways

• Raw scan data is not a deliverable. A single Leica RTC360 setup produces 100 million-plus points, but it only becomes useful after registration, cleaning, classification and modelling — typically 2–10 days of office work per field day.
• Registration accuracy sets the ceiling for everything else. Good cloud-to-cloud or target-based registration holds 2–5 mm across a plant; a poor registration silently bakes 10–20 mm of error into every measurement taken afterwards.
• Georeferencing to GDA2020 / MGA2020 (horizontal) and AHD (vertical) is what makes a point cloud surveyable rather than just a pretty model — without it you cannot tie scans to site control or compare epochs.
• Processing software does not "clean itself". Removing moving plant, dust returns, mixed-pixel edges and vegetation is a deliberate, auditable step that directly affects whether a clash check or deformation analysis is trustworthy.
• Expect realistic field-to-deliverable timelines: a 20–50 setup industrial scan registers in a day, but scan-to-BIM modelling of a congested process plant can run 1–4 weeks and is the largest single cost in most projects.

What a point cloud actually is

A point cloud is a dataset where every measured point carries an X, Y, Z coordinate and usually two extra attributes: reflectance intensity (how strongly the surface returned the laser) and RGB colour (from the scanner's onboard camera). A terrestrial scanner such as a Leica RTC360, Trimble X9 or FARO Focus Premium records this by measuring either the time-of-flight or phase shift of an emitted laser pulse, sweeping a 360° × 270–300° field of view at 1–2 million points per second.

The output of a day in the field is therefore enormous and, in its raw state, almost unusable: dozens of separate scans, each in its own local coordinate frame, full of returns off forklifts, steam, scaffolding and the surveyor's own boots. Point cloud processing is the disciplined sequence of steps that converts that raw capture into a single, clean, correctly oriented, measurable record of the asset. Everything an engineer eventually does with the data — extract a flange face, model a pipe spool, check a 6 mm tolerance, compare this year's kiln shell to last year's — depends entirely on how well that processing was done.

How 3D point cloud processing works: the six-stage workflow

Stage 1: Import and pre-processing

Each scan is imported into the processing platform — commonly Leica Cyclone REGISTER 360, Trimble RealWorks, FARO SCENE, or Autodesk ReCap. At this stage the software applies the scanner's calibration, dual-axis compensator data and any onboard inertial measurements, then builds a quick visual index of every setup. For drone-captured LiDAR (for example a DJI Matrice 350 carrying a Zenmuse L2), this stage also fuses the laser returns with the aircraft's GNSS/IMU trajectory so each point inherits a position. Nothing is measurable yet, but the data is now in a state where it can be aligned.

Tip: Check intensity and trajectory quality before you register anything. A drone flight flown in poor GNSS conditions, or a scan setup that drifted on soft ground, is cheaper to re-fly or re-scan than to fight through processing.

Stage 2: Registration

Registration is the single most important stage and the one most likely to compromise a job. It aligns every individual scan into one common coordinate system using one or more methods:

• Target-based registration — matching surveyed spheres or checkerboard targets that appear in overlapping scans. The most accurate method, standard for tight-tolerance mechanical work.
• Cloud-to-cloud registration — the software matches overlapping geometry algorithmically. Fast and target-free, but it needs 20–30% overlap and enough 3D shape (flat, featureless tunnels defeat it).
• Control-based registration — tying scans to surveyed control points so the whole cloud sits on site coordinates.

Well-executed registration holds 2–5 mm across an entire plant. The danger is that the software will always report a "successful" registration even when it is wrong — so the bundle adjustment residuals and per-setup error must be read and signed off, not assumed.

Tip: On any survey where a measurement will drive fabrication or alignment, use physical targets tied to site control. Cloud-to-cloud alone is fine for visualisation; it is not good enough to cut steel from.

Stage 3: Georeferencing to site datum

A registered cloud is internally consistent but floating in space. Georeferencing rotates and translates it onto the project datum — in Australia that means GDA2020 with MGA2020 grid coordinates horizontally and AHD (Australian Height Datum) vertically, or a defined local mine or plant grid. This is achieved by observing the scan targets, or independent check points, with a total station or GNSS receiver against the existing control network.

Georeferencing is what separates a survey-grade point cloud from a 3D model. It lets you compare a scan taken this shutdown against one from three years ago, set out new steel from the cloud, and hand data to a designer who can trust the coordinates. Skip it and you have a shape, not a survey.

Tip: Verify the realisation of your control. Mixing legacy GDA94 control with GDA2020 scan targets introduces a ~1.5 m shift nationally — a classic, expensive datum mistake.

Stage 4: Cleaning and noise removal

Raw clouds are noisy. Processing removes:

• Moving objects — people, vehicles, swinging loads, which appear as smears or "ghost" points.
• Atmospheric returns — dust, steam, rain and welding fume, which scatter the beam and produce floating points.
• Mixed pixels (edge noise) — where a single pulse straddles two surfaces and returns a false point in between, blurring edges and bolt holes.
• Irrelevant data — temporary scaffolding, lay-down material and adjacent assets outside scope.

This is deliberate, auditable editing. Over-cleaning erases real geometry; under-cleaning leaves artefacts that corrupt a clash check or a 6 mm clearance measurement. On dusty mine sites — a Pilbara crusher station or a Bowen Basin coal handling plant — cleaning is often the longest processing step.

Stage 5: Classification and segmentation

Points are then sorted into meaningful categories — ground, structural steel, piping, equipment, vegetation — using a mix of automated algorithms and manual review. For drone LiDAR over a stockpile or tailings dam, classification separates bare-earth ground returns from vegetation so a true surface model can be built for volume calculation. For a process plant, segmentation isolates the pipe runs and steelwork that will be modelled. Good classification is what makes the later modelling efficient and the deliverable interrogable.

Stage 6: Modelling and deliverable generation

The clean, classified, georeferenced cloud is finally turned into what the client actually asked for:

Scan-to-BIM modelling of a congested plant is labour-intensive — a competent modeller produces a limited volume of intelligent geometry per day — which is why this stage, not the field work, usually dominates the budget.

Accuracy: what processing can and cannot fix

Processing cannot improve the raw measurement accuracy of the scanner — a Leica RTC360 captures roughly 1 mm at 10 m, 2 mm at 100 m, and that is the floor. What processing can do is preserve or destroy that accuracy.

Watch out: The most common failure is trusting the software's automatic registration report. A "passed" registration with high residuals on two setups can put a 15 mm twist into one corner of a plant — invisible on screen, fatal when you fabricate a tie-in spool from it.

For tight mechanical work — kiln shell ovality, crane rail alignment, mill trunnion checks — the realistic, well-processed deliverable accuracy in a real industrial environment is 2–6 mm, not the manufacturer's brochure figure.

Cost considerations

The field day is rarely the expensive part. Processing is.

Indicatively, a simple registered-cloud-only deliverable might sit in the low thousands of AUD; a fully modelled scan-to-BIM of a process unit can run from roughly AUD $10,000 to well over $50,000 depending on the modelling scope. The saving is in avoided fabrication errors and eliminated return site visits.

Common mistakes to avoid

Treating the point cloud as the finished product. A raw cloud is data, not information. Without cleaning and classification it is a liability in a clash check.

Accepting registration on faith. Always read the residuals. Sign off the numbers, not the picture.

Ignoring the datum. Delivering on a local grid when the client needs MGA2020/AHD — or mixing GDA94 and GDA2020 control — is a slow, embarrassing, expensive error to unwind.

Under-specifying the deliverable. "Just scan it" leads to a 60 GB point cloud the client cannot open and a modelling bill they did not expect. Define formats, level of detail and what gets modelled before mobilising.

Frequently asked questions

How long does point cloud processing take?

As a rule of thumb, plan for 2–10 office days per field day. Registration of a 20–50 setup industrial scan is roughly a day; cleaning and classification add days on dusty sites; scan-to-BIM modelling of a congested plant runs 1–4 weeks and is the main schedule driver.

Why register at all — can't the scanner just record one big cloud?

A single setup only sees what is in line of sight. Real assets need many setups from different positions, each in its own local frame. Registration is what stitches them into one consistent cloud; there is no way around it for anything larger than a single object.

How accurate is a processed point cloud in a real plant?

For well-executed industrial work, expect 2–6 mm. Manufacturer specs (1–2 mm) are measured on clean targets at close range in ideal conditions; dust, mixed reflectance, grazing angles and registration overhead all add to that in the field.

What software is used to process point clouds?

Capture-vendor platforms (Leica Cyclone, Trimble RealWorks, FARO SCENE, Autodesk ReCap) handle registration and cleaning; CAD/BIM tools (AutoCAD, Revit, MicroStation) handle modelling. Free viewers such as CloudCompare let clients open and measure deliverables without a licence.

Does processing differ for drone LiDAR versus terrestrial scanning?

The core stages are the same, but drone LiDAR adds trajectory processing (fusing laser returns with the aircraft's GNSS/IMU path) and leans heavily on ground classification for surface models and volumes. Drone operations in Australia must also be flown under CASA Part 101 by a licensed operator.

Talk to us about your scan data

Point cloud processing is where a survey is won or lost. A clean field capture handed to weak processing produces an unreliable deliverable; a disciplined registration, georeferencing and modelling workflow turns the same data into a 2–6 mm record you can fabricate, align and audit against for years. Industrial Spatial Solutions operates Leica scanning hardware and processes data to GDA2020/MGA2020 and AHD across mining, energy and industrial sites Australia-wide.

If you have scan data that needs processing into CAD or BIM — or a shutdown, retrofit or as-built capture coming up — call us on 0407 057 015 to discuss scope, deliverables and a fixed-price quote.