Scanning a Car to Use as a Reference Model

If there's no model of your donor car, you can make one. A phone with a LiDAR sensor, or any phone running photogrammetry, will produce a mesh good enough to design a frame around — provided you treat the output as a rough shape with no trustworthy dimensions until you've proved otherwise.

That proviso is the whole tutorial.

What you get, and what you don't

A scan gives you shape. Surfaces, proportions, where things are relative to each other, roughly.

It does not give you scale you can rely on, flat surfaces that are actually flat, or anything resembling a hard point. Photogrammetry in particular has no inherent scale at all — it reconstructs geometry from parallax and has no idea whether it's looking at a car or a model of a car. LiDAR scans carry a scale and it is routinely off by a percent or two, with drift across a large object.

So: use a scan for packaging volumes and body contours. Do not pull a suspension pickup coordinate off one. Measure those with a tape and a plumb bob like a person.

Scanning well

The capture determines everything downstream, and a bad scan can't be rescued.

• Diffuse, even light. Outdoors on an overcast day is nearly ideal. Direct sun gives you hard shadows that photogrammetry reads as geometry.
• Kill the shine. Glossy paint and glass are the enemy — specular highlights move as you walk, so the algorithm sees a surface that's changing shape. Dulling spray, a dusting of talc, or just accepting you'll lose the glass and bridge it manually.
• Move slowly and overlap heavily. Every frame should share most of its content with the previous one.
• Multiple heights. Walk the car at knee height, waist height and above. Single-height orbits give you a scan that's hollow where you didn't look.
• Put a scale reference in the scene. A metre rule or a printed checkerboard, laid on the floor, in shot throughout. This is the single most useful thing you can do and it takes thirty seconds.
• Don't move the car. Or anything else. The reconstruction assumes a static scene.

For an interior — which is what you want for cage work — scan in sections and expect to do several passes. Interiors are cramped, poorly lit and full of dark trim, which is the hardest combination there is.

Cleaning it up

Raw scans are enormous and full of rubbish. Before importing, in a mesh tool like MeshLab or Blender:

1. Delete the environment. The floor, the walls, the bystander. Everything that isn't the car.
2. Remove floating fragments. Most tools have a "select by connected component" filter — anything not attached to the main body is noise.
3. Decimate hard. A scan can arrive at several million triangles. For a packaging reference you want tens of thousands. Decimating to 5% of the original is normal and you'll struggle to see the difference at the scale you're working.
4. Export as STL or OBJ.

Don't try to make the mesh beautiful. Nobody is going to look at it except you, and only while you have something else selected. Time spent repairing holes in a wheel arch is time not spent designing a frame.

Scaling it, properly

Import it, then prove the scale before you do anything else.

If you put a metre rule in the scene, measure it in the model. Divide, apply the factor, done.

If you didn't, find a dimension on the car you can verify independently and measure that — wheelbase against published specs is the usual choice, and measure it on the real car if you can rather than trusting a spec sheet, since wheelbase varies with trim and nobody is consistent about how they quote it.

Then — and this is the step that matters — check a second, independent dimension. Track width, overall length, the distance between two door handles. If correcting the scale on dimension one doesn't bring dimension two into agreement, the scan has drifted and is not uniformly scaled.

Drift is common on large objects and it isn't fixable with a scale factor. What you do instead is scale to match the region you care about — get the occupant cell right and accept that the nose is 15 mm out — and then stop trusting the far ends of the scan for anything dimensional.

Place it and then verify against reality

Position the mesh against an origin you chose deliberately. Centerline at the front axle is as good a convention as any, and it makes symmetry straightforward.

Then take a handful of real measurements on the actual car with a tape — roof height above the sill, distance from the B-pillar to the seat mounting, width across the floor at two stations — and compare them against the same measurements in the model.

Four or five checks is enough to tell you whether the scan is good to 5 mm or good to 30 mm. Both are usable. Knowing which one you've got is what determines how much clearance you design in.

Designing against a scan

Work with the uncertainty rather than pretending it isn't there.

Hold more clearance than you would against a CAD model. If the scan is good to 10 mm, a 5 mm clearance is a coin toss. Design to 20 mm and stop worrying.

Measure hard points physically. Mounting positions, pickup locations, anything you'll bolt to. Take those off the car with a tape and transfer them as explicit joints, overriding the scan wherever the two disagree. A scan is a shape, and when a shape argues with a tape measure the tape wins.

Confirm the critical fits before cutting steel. Make a cardboard or welding-rod mock-up of the one or two members that pass closest to the bodywork and offer it up to the real car. Fifteen minutes with a coat hanger has saved a great many main hoops.