Structured light scanning of large objects: field of view and accuracy

A train front 2.6 m wide and 2.3 m high will not travel to a measuring room, will not fit on a CMM, and point-to-point hand measurement cannot tell you how it conforms to drawing. Structured light scanning answers that by moving the instrument instead of the part: the same projection head that measures a phone covers a vehicle front once you widen its field of view, and you stitch the captures into one model. This guide sizes that field of view, prices the widening in point spacing and depth accuracy, and covers what decides success at this scale — calibration board size, ambient light and stillness, and alignment targets. The numbers are planning values from fringe-projection geometry plus one documented rail job; the quoted accuracy of the head you buy always overrules the planning rule.

Can one structured light scan cover a whole 2.6 m train front?

No. A single capture from a portable fringe-projection head covers roughly 0.4 to 1.2 m on a side, so a 2.6 m by 2.3 m front needs about 16 scans at 30% overlap, and 20 to 25 once you count returns, gutters and roof edges.

Snapshot structured light triangulates a whole area in one exposure: a projector throws fringe patterns onto the surface, one or two cameras photograph the deformed fringes, and phase unwrapping gives one depth point per camera pixel in 0.1 to 2 s. The part never moves; the head travels on a tripod, jib or scaffold. One rail front was captured top-down from 5 m up with the head slung from a jib, at a weight one person can handle.

The field of view obeys the ordinary lens equation — sensor width, focal length, stand-off — which is why swapping lenses and stand-off turns a phone-scale head into a vehicle-scale head without touching the sensor. When your 2D inspection stage hits the same wall, the cousin of this problem is locating a large part when the camera field of view is too small. That is the shape of the job, and the instrument class you will find among our 3D sensors:

What does widening the field of view cost in accuracy?

Point spacing grows in direct proportion: a 1,280-pixel camera row gives 0.23 mm per point at a 300 mm field of view and 0.94 mm at 1,200 mm, and depth noise rises with it.

The first two lines are geometry, exact. The third is the planning rule for fringe projection: quoted depth accuracy on these systems tends to land near one ten-thousandth of the field of view, so use it to plan the job and let the quoted accuracy of the head settle the argument. At 1,200 mm that rule gives 1,200 / 10,000 = 0.12 mm per scan.

Worked example, one head, two configurations. The cameras are 1,280 x 960 with a 7.2 mm sensor width:

The arithmetic, wide mode: FOV = 7.2 mm x 2,000 mm / 12 mm = 1,200 mm. Point spacing = 1,200 / 1,280 = 0.94 mm. Height follows the sensor aspect: 900 mm. Scan centers step by 70% of the field, 840 mm across and 630 mm up, so 2,600 / 840 = 3.1, round up to 4 columns, and 2,300 / 630 = 3.7, round up to 4 rows: 16 captures. In detail mode the same part takes 2,600 / 210 = 12.4, so 13 columns, and 2,300 / 157.5 = 14.6, so 15 rows — about 200 captures, which is why nobody scans a train front at phone resolution.

The wide lens pays one bonus: depth of field. Magnification falls from 7.2 / 300 = 0.024 to 7.2 / 1,200 = 0.006, and geometric depth of field grows with the inverse square of magnification at fixed aperture — about sixteen times deeper — so surfaces well in front of and behind the sweet spot stay sharp. That reach matters on curved vehicle panels.

The spacing arithmetic is the same one behind depth camera pixel size estimated from field of view and resolution, and it is why changing a 3D camera's depth resolution and field of view at capture time deserves a deliberate choice rather than a default. Run your own sensor and stand-off numbers before you commit to a lens set:

The small end of the same trade sits in the lighting catalog, where a pattern projector's field is the part itself: structured light panels are small-field tools by nature, because that is where a fine pattern pitch pays. The LS-SPC-0103 is a 200 x 150 mm structured light source.

How big should the calibration board be?

Match the board to the field of view, not to the part: about half the FOV width, which at 1,200 mm means a 720 x 540 mm board of 60 mm squares on a 12 by 9 grid, shot in 30 to 40 poses.

The sizing comes straight off the grid: 60 mm squares x 12 = 720 mm wide, x 9 = 540 mm high. Smaller squares waste sensor pixels on the board and bias the phase-to-metric fit; a board far under half the field under-constrains the lens model at the edges, exactly where wide lenses distort most.

Large accurate boards are genuinely hard to make: flat to a small fraction of one square across the whole span, opaque, matte, dimensionally stable. If you scan at this scale often, build one properly — the construction methods in building a large calibration grid for a 1 m field of view carry across to 720 mm. Store it flat, never on edge.

Calibrate per configuration: each lens and stand-off pair is a different camera-plus-projector geometry and needs its own board session. Thirty to forty poses spread through the working volume cut distortion residuals and processing time together; fewer than about twenty shows up later as bowl-shaped error across the stitched model. When a session that worked refuses to repeat, the failure pattern in structured light calibration failing on the second shot is the usual map.

How much ambient light can the scan tolerate?

Very little: direct sun is about 100,000 lux and no portable projector out-illuminates it, so outdoors means night or full shade, while a 300 to 500 lux shop floor is workable with local shading.

Daylight sets the plan: full sun 100,000 lux, overcast 1,000 to 10,000 lux, and a portable projector typically lands only a few hundred lux on a part this size. Overcast daylight buries the fringes; a dim aisle with the overhead row above the part switched off usually does not.

The field check is contrast, not a meter. Open the live view with the pattern running and compare a fringe valley with a fringe peak: when the valley sits below roughly a quarter of the peak, ambient is low enough that phase noise stays near the dark-shot floor. When it does not, shroud, shade or wait. It is the same brightness-versus-coverage balance as large field of view lighting, with the projector standing in for the LEDs.

Two heat sources count as light problems too. Duct the projector exhaust so it never blows warm air across the cameras' sight lines, and give a head that just came out of a cold van time to stabilize. Both show up as shimmer in the zoomed live view before they show up in the data.

How still do the scanner and the part have to be?

Budget motion below about a fifth of the point spacing during each capture — roughly 0.2 mm at 0.94 mm spacing — so a breeze loading a 5 m jib, or a part still settling, will smear the fringes before anything else fails.

At a typical 0.5 s capture, that budget is a velocity limit of 0.4 mm/s, which light wind on a long boom exceeds easily. Let the jib settle after every repositioning; three to five seconds of dead hang is usually enough. The discipline is the one behind 3D camera mounting stillness during capture — the instrument has to behave like survey equipment, not a handheld camera.

How do you align and stitch the scans into one model?

Give every scan at least 30% overlap with what is already captured and at least three non-collinear reference features inside that overlap — on a smooth front you add them yourself, and 100 to 250 mm matte targets are the right scale.

Natural geometry comes first: window rubbers, panel swages, fasteners and dents lock frames together at no cost. Where the surface is featureless, add reference objects — on one rail front, half styrofoam bowls and plates of roughly 150 to 250 mm did the job, light enough to tape on and big enough to constrain a 1.2 m frame from any angle. Three targets minimum, never in a straight line, placed so every planned capture sees at least three of its neighbor's targets.

Errors accumulate the way measurement errors do. Pair-to-pair alignment lands near the per-scan depth figure, 0.1 to 0.15 mm; across a 16- to 25-capture chain with targets, global error holds near 0.5 to 1 mm. Drop the targets and trust the surface alone, and drift of several millimeters across the model is the normal outcome, not the exception.

Run two closure checks before packing up. Re-capture the first scan position last and compare it to the merged model — the mismatch on shared targets is your global error. Then read the neighbor deviation map: healthy overlap shows green through the shared band, with red only on grazing surfaces. At 0.94 mm point spacing the final cloud still resolves the tape holding the markers on, which is the level of detail that keeps a scan of this size worth re-measuring years later. From there, export follows the ordinary route of taking one capture to a structured point cloud as XYZ and PNG.

When is structured light the wrong tool for a large object?

Move off snapshot structured light when the surface runs continuously past a fixed point, when stand-off must pass about 3 m, or when the tolerance applies across the whole 2.6 m rather than locally.

Continuous product — extrusion, strip, rail stock passing a station — wants a line-scan 2.5D head that maps in motion, frame after frame, with the conveyor as the second axis. The MV-HLVS-16K100-L60 is that class of instrument: 16K resolution, a 120 kHz sampling rate and a 600 mm span in one head. Coarse volume — pallets, bins, load plans — wants depth-per-pixel speed, not sub-millimeter points; a ToF sensing camera such as the MV-DT01WMU fills that role, with the known limit that small distant objects lose detail in a ToF point cloud. Two heads cover those two detours:

The third case is tolerance. If the drawing demands 0.1 mm over the full 2.6 m — profile against a master model, not local dent depth — a stitched portable project will not certify it. Add a global frame: photogrammetry or a laser tracker sets the network while the scanner fills in the detail, or the part goes back to large-capacity metrology. Sub-0.1 mm work is structured light's home ground at small fields, as why 0.1 mm on a foot-long part needs structured light shows from the other end of the scale.

What should you check before the rig goes on site?

Walk this list before the rig leaves the shop: field-of-view mode chosen and calibrated for, targets cut and mapped, the light plan agreed, and a closure re-capture planned as the last shot of the night.

Optics and calibration

• Lens set and stand-off locked before calibration; a mode changed on site is a mode recalibrated on site.
• Depth-of-field reach checked against the part's curvature — the wide lens's deeper focus is what keeps panel flanks usable.

Site

• Night window or full shade booked; overcast daylight is not workable.
• Projector exhaust ducted away from the camera sight lines; cold equipment given time to stabilize.
• Boom settle rhythm agreed, with a wind limit that says stop rather than hope.

Targets

• Targets cut, matte, and placed so every capture sees three non-collinear references.
• Placement photographed, so they can be found on the part a year later.

Verification

• First scan position marked for the closing re-capture.
• Deviation-map review written into the run sheet and judged on site.

The decisions, then, per case: a one-off 2.6 m part held to about 1 mm globally is a wide-field structured light job done in one evening with targets and a night window. The same part held to 0.1 mm across its full extent needs a global frame around the scanner, not a scanner swap. A surface that never stops needs line-scan 2.5D, and mere volume needs ToF. Pick the head by field of view and accuracy in the 3D sensor finder, size the whole rig in the system configurator, or send us the part size and tolerance and we will lay out the scan plan with you.