Point cloud to BIM with AI: where scan to BIM works and where it breaks

· Postlight Research

Scan to BIM turns a laser survey of an existing building into an editable model that designers and engineers can work in. Machine learning removes much of the manual tracing, but not all of it, and the parts it leaves behind are predictable. This note covers the capture choice, how a system separates the building from the clutter inside it, where automated modelling fails, and how to check the result against the scan.

Capture: terrestrial scanning or mobile LiDAR

Two families of equipment produce the point clouds that scan to BIM starts from.

A terrestrial laser scanner sits on a tripod at a series of stations and measures millions of points from each. Stations are registered to each other through overlapping geometry or targets, and tied to survey control when the model has to sit in project coordinates. The result is dense and accurate, and slow to collect: one floor of a large building can take a day of station moves.

A mobile scanner, handheld, backpack or trolley mounted, records continuously as the operator walks, using SLAM to estimate its own path. It covers the same floor in a fraction of the time. The cost is drift: small errors in the estimated path add up over a long walk, which bends long corridors and opens gaps where loops should close. Closing loops, walking back through known areas and tying in a few control points keep drift in check.

Illustrative: the usual trade-offs. Check them against the specific instrument and procedure.

Terrestrial laser scanning Mobile LiDAR
Capture speed Slow, station by station Fast, one continuous walk
Accuracy Highest; suits tight tolerance work Lower, and falls with path length without control
Tight spaces Needs many stations Follows the operator into small rooms and risers
Best use Structural deviation, facades, plant rooms Space planning, fit-out, fast surveys of large floors

The right choice depends on the accuracy the model has to meet. The USIBD Level of Accuracy specification gives a shared language for it: it bands accuracy from LOA10 (coarser than 50 mm) to LOA50 (within 1 mm) at a 95 percent confidence level, and it keeps the accuracy of the measured data apart from the accuracy of the modelled elements. A model cannot be more accurate than the cloud it came from, so set the target before choosing the scanner.

Separating the building from what is in it

An occupied or active building is full of things that are not the building: furniture, stored materials, scaffolding, people, vehicles, temporary partitions. A point cloud records all of them as faithfully as the walls.

Automated scan to BIM works in stages:

  1. Clean. Remove noise, the mixed points that form at edges, and points from moving objects, which show up as ghost trails in mobile scans.
  2. Segment. Classify points as wall, floor, ceiling, column, door, window or clutter. Point cloud networks trained on labelled building scans do this point by point; simple geometric rules, such as large vertical or horizontal planes, still do much of the work for walls and slabs.
  3. Fit. Fit parametric elements to the segmented points: a wall becomes an axis, a thickness and a height; a slab becomes a boundary and a depth; a column becomes a profile and a height; doors and windows become openings hosted in their walls.
  4. Assemble. Join walls at corners, attach them to slabs, place elements in storeys and spaces, and export to IFC4 or a Revit model with real element types rather than meshes.

The difference between a usable model and a mesh with labels lies in steps 3 and 4. Engineers need a wall they can move, a slab they can cut and a door schedule they can read. Fitting parametric elements, not triangulated surfaces, is what makes the model editable.

Where automated scan to BIM breaks

Automation fails in a small number of recurring ways, and a good system reports them instead of hiding them under a confident model.

Occluded ceiling voids. Above a suspended ceiling the scanner sees nothing unless tiles are lifted. Services, beams and the slab soffit there are guessed or left out. The honest output marks the void as not captured.

Reflective and transparent surfaces. Glass, polished stone, mirrors and stainless steel return false points behind the surface, or no points at all. Curtain walls and glazed partitions often need manual modelling or a separate capture with targets.

Unlabelled MEP runs. A scan shows a pipe's position and diameter well. It cannot tell a chilled water flow from its return, or a gas line from a sprinkler main, without labels, colour coding or drawings. The geometry is recoverable; the system it belongs to often is not.

Non-orthogonal and altered geometry. Recognition models learn from typical buildings: straight walls, square corners, repeated floor plates. Older buildings with walls out of plumb, curved bays and decades of alterations break those assumptions. A wall that leans 20 mm over its height should be modelled as it stands, not snapped to vertical, if new work will be built against it.

Clutter that looks permanent. Shelving fixed to a wall, a row of tall cabinets or a temporary partition can be segmented as wall. Context from neighbouring spaces and from photographs helps; a human check of doubtful elements still pays for itself.

Checking the model against the cloud

A scan to BIM deliverable should come with proof that it matches its source. The standard method is a cloud to model deviation check:

  1. Export the model as surfaces in the same coordinate system as the point cloud.
  2. For every point, compute the signed distance to the nearest model surface.
  3. Colour the cloud by that distance, with one band for in tolerance and separate colours for points in front of and behind the model.
  4. Report statistics per element type: the mean offset, which reveals a systematic shift, and the spread, which reveals a poor fit.
  5. List the elements where a large share of their points fall outside tolerance, and review those by hand.

Illustrative: a tolerance check by element type, with limits set by the project's accuracy target.

Element type Typical check What a failure usually means
Walls Points between the wall faces Wrong thickness, or a lean modelled as vertical
Slabs Offsets of soffit and top surface A sagging slab or a missed change of level
Columns Centre position against the grid A column off grid or a misread section
Openings Jamb and head positions A door or window moved during an alteration

Run the check on the full cloud, not a cleaned subset, and keep the uncovered areas visible. A deviation report that covers only the places where the model fits is not a check.

Using the result

An automated scan to BIM model is a starting point that a team trusts in proportion to the evidence behind it: the capture method, the registration report, the deviation check and a plain list of what was not captured. Delivered that way, it saves the weeks of manual tracing that make surveys of existing buildings expensive, and leaves engineers free to spend their time on the doubtful parts.

Scan to BIM for existing buildings is part of Postlight's site twin work: see scan to BIM for existing buildings. For new construction, where the design model exists and the job is keeping it current, read BIM progress monitoring; for design coordination, read BIM clash detection and triage.