Intraoral Scan Point Cloud Registration: From Gingival Interference and Curvature Sampling to Improved ICP

Based on Lu Pingyue, "Research on Point Cloud Registration Technology for Intraoral Scanning Systems," Master's Thesis, University of Electronic Science and Technology of China, 2023.

This article reorganizes the thesis around a clear problem–method–experiment–engineering–limitations structure. It focuses on reusable technical ideas for better dental digital models and does not evaluate any commercial device or solution.

Why Point Cloud Registration Matters More Than Single-Frame Capture

An intraoral scanner captures only a small local surface of the dentition in each frame. Building a complete digital oral model requires transforming dozens or even hundreds of these local point clouds into a common coordinate system.

Single-frame reconstruction errors mainly affect local geometry. Registration errors, however, propagate along the scanning path. Wrong correspondences can produce double images, gaps, holes, cusp shifts, or progressive bending of the dental arch.

The core research question is:
How can the inherent scene characteristics of intraoral scan point clouds be used to improve the classic registration pipeline so that the algorithm focuses more on the tooth region while controlling the computational cost of preprocessing and fine registration?

The thesis does not invent a new framework. It improves the classic "preprocessing → coarse registration → fine registration → meshing" pipeline with four practical modifications tailored to intraoral scanning:

Point-cloud segmentation + camera coordinates and viewpoint distance to remove noise and part of the gingival points

One-dimensional curvature histogram for feature-preserving down-sampling

Viewpoint-distance-based error weighting inside point-to-plane ICP

Combination of Euclidean distance and FPFH feature distance to filter suspicious correspondences

Thesis and Project Overview

Experiments used a structured-light prototype (camera 1920×1080 @ 60 fps, projector 1920×1080 @ 168 fps). Test platform: AMD Ryzen 7 5800H, 16 GB RAM, NVIDIA RTX 3050. Results are module-level validations on experimental point clouds, not full real-time product benchmarks.

The Real Problems the Work Addresses

1. Gingival points can "register better" but make teeth register worse
Intraoral point clouds contain teeth, gingiva, and noise. Larger, smoother gingival regions can dominate least-squares registration. Clinically, however, cusps, fossae, and margin lines matter most. Preferring large gingival areas can misalign occlusal surfaces. Distant regions also suffer more from calibration error, reconstruction noise, and occlusion. The thesis uses a practical prior: in most local frames the teeth are closer to the camera viewpoint than the posterior gingiva. Denoising and ICP weighting are built around this assumption.

2. Dense clouds slow registration; uniform down-sampling loses dental features
Typical experimental clouds contain 22 000–28 000 points. Direct normal estimation, FPFH, nearest-neighbor search, and iterative registration become expensive. Voxel filtering is fast but can undersample grooves and sharp edges. Normal-space sampling preserves high-curvature regions but requires many bins in 3-D normal space. Curvature-interval sampling reduces the histogram to one dimension while still favoring feature-rich areas.

3. Classic ICP accepts geometrically close but feature-dissimilar correspondences
On repetitive dental surfaces, small Euclidean distance does not guarantee matching local geometry. The thesis therefore adds FPFH feature distance as a second filter during fine registration.

Method 1: Segmentation + Camera Coordinates to Extract the Tooth Region

Pipeline:
radius outlier removal → normal & curvature estimation → region growing → camera-coordinate constraints → small-region deletion → keep nearest regions

Radius outlier removal removes isolated reconstruction noise.

Weighted local-plane fitting estimates normals (oriented toward the known camera) and curvature.

Region growing (relaxed normal-angle constraints) exploits natural breaks caused by interdental spaces, shadows, and occlusion.

Points are transformed into the camera frame; regions outside an effective 3-D range are discarded.

Regions smaller than 10 % of the largest remaining region are deleted; the two closest regions to the viewpoint are kept (to avoid discarding teeth split by shadows).

On four test clouds, denoising took 255–335 ms (dominated by region growing and normal/curvature estimation). Strict 0.2 mm registration recall improved by approximately 20.3 percentage points on average. Limitation: when teeth and gingiva form a single connected component, the method cannot separate them by connectivity alone.

Method 2: Curvature-Interval Sampling

Instead of simply keeping the highest-curvature points (which may lie outside the overlap region), the curvature range is divided into bins and points are sampled starting from high-curvature bins while still retaining some mid- and low-curvature points. This preserves features yet keeps enough common support for coarse registration.

Compared with normal-space sampling, one-dimensional curvature bins scale better at larger target sizes:

Final registration accuracy of the three methods was similar; the main value of curvature sampling is lower bin-maintenance cost while retaining a feature bias. Voxel filtering remains the fastest option when pure speed is required.

Method 3: FPFH + SAC-IA for Coarse Registration

Classic two-stage strategy:

Coarse: FPFH (33-dimensional) + SAC-IA

Fine: point-to-plane ICP

Clouds were down-sampled to 5 000 points. SAC-IA with 100 iterations (chosen as a practical trade-off) averaged ~2.2 s. Mutual-correspondence constraints improved accuracy; early termination reduced time but lowered stability. SAC-IA is therefore better suited to initialization, recovery after tracking loss, or offline use than to every frame of high-frame-rate scanning.

Method 4: Making ICP Prefer Reliable Tooth Correspondences

Viewpoint-distance weighting
Target points closer to the camera receive higher weight in the point-to-plane error. The goal is better alignment of the (usually more reliable) near-tooth region rather than minimal overall RMSE.

Euclidean + FPFH fusion distance
For each ICP correspondence both distances are computed, normalized, and combined. Only the lower-fusion-distance fraction (stable in the 0.4–0.8 range) participates in the transformation estimation-similar in spirit to Trimmed ICP but using feature information as well.

On six test clouds the combined improvements raised 0.2 mm recall by ~7.1 percentage points on average. RMSE and 0.6 mm recall did not improve on every set; in some cases the proportion of high-precision correspondences increased while overall RMSE rose. The method therefore optimizes the share of high-accuracy alignments rather than every metric simultaneously.

Runtime of plain point-to-plane ICP was 27–45 ms; with fusion constraints it rose to 51–126 ms.

Software Prototype Scope

The prototype implements:

Preprocessing (outlier removal, normals/curvature, region growing, coordinate filtering, curvature sampling)

Registration (FPFH, SAC-IA, viewpoint-weighted ICP, fusion-distance constraints)

Meshing (voxel filtering, moving least squares, ball-pivoting triangulation, non-manifold repair, hole filling, edge optimization, face-normal correction)

Stack: C++11, Visual Studio 2017, PCL 1.8.1, OpenCV 4.3.0, VTK 8.0.0, Qt 5.12.1, plus camera and DLP SDKs.

Demonstrated modules work end-to-end on a few frames, but continuous frame rate, end-to-end latency, long-sequence drift, tracking failure/recovery rates, full-arch absolute accuracy, and comparison against clinical reference models are not reported. The software is an algorithm-validation prototype, not a clinically verified real-time system.

Value and Limitations

Reusable strengths

Methods are tightly coupled to real intraoral priors (teeth closer to camera, large gingival areas, repetitive local geometry).

Improvements stay inside well-understood modules (region growing, curvature, FPFH, SAC-IA, ICP) that are easy to implement and debug in C++/PCL.

Dual-threshold recall (0.6 mm and 0.2 mm) distinguishes overall fit from high-precision local fit.

Caveats

Small sample size (mainly 4–6 clouds).

Fixed geometric priors fail when teeth and gingiva are connected or scanning angles change strongly.

SAC-IA cost (~2.2 s at 5 000 points / 100 iterations) precludes per-frame use.

No global optimization, loop closure, or pose-graph treatment of long sequences.

Limited comparative statistics and no external clinical ground-truth models.

Practical Engineering Takeaways for Intraoral Scanning Systems

Not every point should have equal weight. Combine viewpoint distance, reconstruction confidence, normal stability, curvature, reflection masks, and (when available) semantic labels into a multi-factor reliability model.

Keep expensive global registration out of the normal tracking loop. Use previous-pose prediction + local ICP for routine frames; expand search or switch to keyframes when quality drops; invoke FPFH/SAC-IA/FGR or learned descriptors only for recovery; apply sliding-window / loop-closure / pose-graph optimization for long sequences.

Move evaluation from pairwise error to full-arch clinical metrics. Report trueness and precision on single teeth, half-arch and full-arch, cumulative drift, critical-region (cusps, margins, contact areas) error, inter-operator and inter-path repeatability, robustness to moisture/reflection/restorations, and end-to-end latency/memory.

Summary

The thesis constructs a complete, interpretable pipeline for intraoral point-cloud registration: remove gingival and noise interference with segmentation and camera priors, compress the cloud with curvature-interval sampling, obtain an initial pose with FPFH + SAC-IA, and refine with viewpoint-weighted, feature-distance-constrained point-to-plane ICP. The improvements raise the proportion of high-precision tooth correspondences on the tested data while highlighting the remaining challenges of connected gingiva, coarse-registration cost, and limited evaluation scale.

From a modern engineering perspective the work serves as a clear traditional baseline. Its lasting lesson is that registration quality in intraoral scanning depends heavily on correctly defining which points are trustworthy and which correspondences deserve to influence the optimization.

Source note
This article is based on Lu Pingyue's 2023 master's thesis. Experimental numbers, device parameters, and conclusions follow the original text; statements about percentage-point gains, real-time suitability, and engineering scope are technical analyses derived from those data.

Looking for reliable intraoral scanning that delivers accurate full-arch digital models?
Modern systems such as the Aident AI-30 (≤10 μm accuracy, ultra-lightweight design, powder-free true-color scanning, open STL/PLY/OBJ export) are engineered precisely to minimize the registration and cumulative-error issues discussed above and integrate seamlessly into Scan → Design → Print workflows.

Explore the AI-30 and complete digital solutions at https://www.aident3d.com/ or contact us for a demonstration and ROI comparison.