Point Cloud Registration Alignment

Two scans of one room start misaligned. Registration finds the rotation and shift that snap them together for mapping and tracking.

Coarse matches seed a fine alignment that snaps two scans together

Why Does This Exist?

Lidar frames, room scans, and CAD models live in different coordinates until aligned. Registration exists as the snap step behind mapping, visual SLAM loop repair, and object pose: find correspondences, solve $T = [R|t]$ , verify on overlap. Without it scans stay as floating shells.

Local refinement is ICP; learned global matching is Deep Global Registration; image analog is image registration.

Think of It Like This

Aligning two transparencies of stars

Lay two star transparencies offset on a projector and slide one until constellations overlap. A rough corner match gets close; nudging individual stars locks it.

Coarse matches are the corner slide and ICP is the nudge. The analogy stops at outliers: stars rarely lie while scans carry moving people that must be ignored.

How It Actually Works

Coarse stage extracts descriptors such as FPFH or learned FCGF, matches them, and runs RANSAC for a rough $R$ and $t$ that survives up to $70$ percent outliers. Fine stage runs point-to-plane ICP on the overlap region. Success needs overlap above about $30$ percent; below that symmetry and repetition dominate. Error reports translation in centimeters plus rotation in degrees on the overlapping points only.

Worked numbers

Two room scans share $60$ percent overlap. Coarse RANSAC from $500$ matches keeps $120$ inliers and lands within $0.3$ m and $5$ degrees. ICP refines to $0.04$ m and $0.8$ degrees mean error on overlap. A $10$ percent-overlap pair from opposite doors fails with $2$ m error despite low residuals on the sliver.

Code

1def apply_rigid(pts: list[tuple[float, float]], angle_deg: float, tx: float) -> list[tuple[float, float]]:2    import math3    a = math.radians(angle_deg)  # 2D stand-in for 3D rigid motion4    return [(x * math.cos(a) - y * math.sin(a) + tx, x * math.sin(a) + y * math.cos(a)) for x, y in pts]5
6
7print(apply_rigid([(1.0, 0.0)], 90.0, 0.5))  # -> [(0.5, 1.0)]

Watch Out For

Registering near-duplicate hallways

Symmetric corridors give many equally good snaps offset by one door spacing. The symptom is low residual with wrong room assignment. Verify with a second modality such as image overlap or free-space checks before merging maps.

The Quick Version

Registration solves the rigid $R$ and $t$ aligning two scans.
Coarse descriptor matching plus RANSAC survives heavy outliers.
Fine ICP polishes on the overlap region only.
Overlap above $30$ percent is the practical minimum.
Reference rooms land under $5$ cm and $1$ degree with $60$ percent overlap.
Symmetry needs extra verification beyond low residuals.