Corner Anchor

Every sensor arrives in its own frame

An RTK drone gives absolute position with a bias of a few centimetres. A ground 360 capture gives excellent density and drifting scale. A mobile lidar gives a beautiful cloud in an internal SLAM frame nobody measured. All three describe the same building — and none of them agrees with the others.

The usual answer is to stitch afterwards: ICP between clouds, manual point picking, coarse alignment then refinement. It works, sometimes. The problem is not accuracy — it is that the result cannot be repeated and cannot be documented. There is no number to show the client, no residual to inspect, and if someone rescans in six months nothing ties the two surveys together.

This method inverts the order. Instead of matching clouds after the fact, you place one rigid body whose geometry is known in advance, and every sensor that sees it is tied to it — and through it, to each other.

The corner is the intersection of three lines

This is the heart of the method. An external building corner is not a point somebody marked — it is a point the world has already defined, where three real lines meet.

• A vertical line — where facade A intersects facade B.
• The first base line — where facade A intersects the slab.
• The second base line — where facade B intersects the slab.

Three lines, one point. It needs no paint, no peg and no marking — it is implied by three surfaces the builder already poured. And more importantly: that exact point exists as a vertex in the IFC model. Anchor to it and the survey and the model speak the same language without a separate control network.

The fixture locates on the corner by 3-2-1 kinematic constraint: three contacts on the slab, two on facade A, one on facade B. Six contacts for six degrees of freedom — no more, no fewer. No legs, no levelling, no shims. Set it on the corner and it is located. The CAD origin (0, 0, 0) is the building corner itself.

Not one scale bar — twenty-one

Seven markers produce 21 independent distances, from 0.4752 m to 1.5165 m — a 3.19× spread of scales, every one of them known in CAD.

The difference between one scale bar and 21 is the difference between belief and measurement. With a single bar, a bad observation quietly rescales the model. With 21 independent distances the system is over-determined: an outlier shows up as a residual instead of as a scale error. That residual is something you can put in front of a client.

Three points are enough to fix a rigid six-degree-of-freedom transform. Seven give more than double the redundancy — and therefore robustness: if one marker is obscured, dirty, or falls outside the frame, the solution is still fixed and still checked.

And from there, georeference

Occupy the building corner once with RTK. The moment that point is known in world coordinates, all seven markers are known with it — because their offsets from it are fixed in CAD to the micron. And everything registered to those markers — drone, ground capture, lidar — comes along into the same frame.

The export is a CSV in metres at micron precision, with a 1 mm accuracy field per axis. It loads straight into Metashape as Reference.

Detection is automatic, and the labels bind themselves

Detect Markers with CircularTarget12bit finds the targets and decodes their IDs. The CSV is imported and binds by label. No manual point picking, at any stage.

For that to actually work, the printed rings have to be legal codes. That is less obvious than it sounds: of the 4096 possible 12-bit ring patterns, only 266 decode at all, and they collapse to exactly 161 IDs — 1 through 161. An illegal pattern prints perfectly and is never detected, with no warning of any kind.

The detector also has a size window: the inner dot radius has to land roughly between 5 and 40 pixels. Too large fails exactly as reliably as too small — worth knowing if you photograph a big target from very close.

These targets were sized around a 15 pixel threshold on the inner dot. With a 123.8 mm dot and a Mavic 3E wide lens, that gives automatic detection out to 31 m.

What you do on site

1. Set it on the corner

   The six kinematic contacts seat themselves — three on the slab, two on one facade, one on the other. No levelling, no adjustment.

2. Occupy the corner with RTK

   One measurement of the building corner. From there all seven markers are known in world coordinates.

3. Fly, scan, capture

   Drone at 40° elevation or above; ground station between 2 and 6 m; laser scanner at 2 m if you want automatic detection in the cloud as well.

4. Detect Markers

   CircularTarget12bit. IDs decode themselves and become labels.

5. Import Reference

   Load the CSV. Binding is by label — not a single point picked by hand.

6. Check the residuals

   The 21 distances leave the system over-determined. Those residuals are the evidence you can show a client.

What the method does not do

Every claim here was measured. These are the things that were measured and came back negative, and they belong in the picture as much as the rest.

• Airborne lidar will not decode the rings. Not at the densities current DJI payloads produce. For airborne registration, use the plane geometry, not the targets.
• The drone needs at least 30° elevation at 20 m, and 40° for comfortable margin. A very flat approach sees only two markers.
• Range is 31 m, not 100. The fixture was deliberately scaled down from a 100 m version because the weight there was unreasonable. From 100 m the targets are too small.
• It needs a genuine 90° external corner. The kinematic location assumes two perpendicular facades and a slab. A rounded, clad or out-of-square corner needs a different design.
• 7% of sun directions produce a shadow edge on one marker or another. It does not break the solution — at least four always survive — but it is not zero.

One sheet, two people, a car boot

Zero waste on the sheet, one part repeated seven times, and it goes through an ordinary house door. Building four fixtures for the four corners of a structure consumes 28 of the 161 available IDs — still nowhere near the ceiling.