Is NDT only for three-dimensional data?

No. Two-dimensional and three-dimensional forms exist, although many practical tutorials focus on 3D point clouds.

Does it require a neural network?

No. Classical NDT registration uses statistical modeling and numerical optimization.

Does convergence guarantee correct alignment?

No. It indicates a stopping condition was reached, and independent geometric validation remains necessary.

AI Glossary / Computer vision and multimodal AI / Intermediate

Normal distributions transform

Definition

The normal distributions transform, or NDT, represents local groups of spatial points with Gaussian distributions and can use that representation to align point clouds. Instead of relying only on explicit nearest-point pairs, an NDT registration method adjusts a transformation so points from one scan fit the statistical structure represented by another scan.

What it means

A point cloud is a collection of measured locations, often produced by a depth sensor or laser scanner. Two scans of the same room may contain many corresponding surfaces but express their points in different coordinate systems.

NDT summarizes portions of a reference cloud using a mean location and covariance. The mean describes where a local group is centered, while the covariance describes its spread and directional shape.

The word normal refers to normal, or Gaussian, probability distributions. It does not mean that the algorithm simply aligns surface-normal vectors. NDT can describe local surface geometry statistically without storing a direct point-to-point match for every source point.

How it works

Divide the target cloud's space into cells or voxels and estimate a distribution from the points in each sufficiently populated cell. Cell size matters: very large cells can blur important structure, while very small cells may have too little information for stable statistics.

Supply the source cloud and an initial estimate of its transformation into the target coordinate system. Transform source points and evaluate their agreement with the target distributions according to the particular NDT objective.

An optimizer changes translation and rotation to improve that agreement. The rigid-registration version preserves the source cloud's shape while changing its pose. Practical implementations include numerical safeguards, stopping rules, and parameter choices that must match the data scale.

After optimization, validate overlap and geometry independently. A convergence flag indicates that an algorithm met its stopping condition; it does not prove that the scans were aligned to the correct location or that the scene remained unchanged.

Three practical examples

1. Mobile robotics: A robot aligns a new laser scan with a local map as part of estimating its position.

2. Indoor surveying: Several overlapping room scans are brought into a common coordinate frame before producing a combined spatial record.

3. Inspection research: A depth scan of a rigid object is aligned with a reference before examining residual geometric differences.

Think of it this way

Think of replacing a dense crowd of map pins with a set of soft ellipsoidal clusters, then moving another set of pins until it fits those clusters. The ellipsoids summarize both location and direction of spread.

The analogy is approximate. Actual NDT scoring uses mathematical distributions and a numerical objective, and the cells may describe surfaces rather than separate physical objects. A cluster summary can also lose detail that matters for precise registration.

Worked example

Consider a two-dimensional teaching cell with mean (0,0) and covariance having diagonal entries 4 and 1, with zero cross-covariance. Its spread is larger horizontally than vertically.

A point at (2,0) has squared Mahalanobis distance 2²/4 + 0²/1 = 1. A point at (0,2) has distance 0²/4 + 2²/1 = 4. Both are two ordinary distance units from the mean, but the vertical displacement is less consistent with the cell's distribution.

Ignoring the common normalization factor, their Gaussian density factors are exp(-1/2), approximately 0.607, and exp(-4/2), approximately 0.135. This illustrates how local directional structure affects agreement.

A full registration algorithm evaluates many transformed points and optimizes a defined objective. The example is not a complete NDT score and should not be used to infer a probability that an entire alignment is correct.

Illustrative case study

Fictional case: A facilities team combines scans of a long corridor. Its NDT alignment sometimes shifts a scan by one repeated doorway spacing while still producing a plausible local fit.

The team improves the initial pose estimate and evaluates alignment against independently surveyed control points. It also tests cell sizes appropriate to the corridor's dimensions rather than copying parameters from a small-object example.

When people moving through the scene create inconsistent geometry, the team examines filtering and overlap assumptions. It treats the optimized score as one piece of evidence and keeps a review path for repeated structures or insufficient overlap.

Common misconceptions and mistakes

Do not confuse normal distributions with surface normals or with a method for making arbitrary data normally distributed.

Do not assume NDT is a guaranteed global search. Initial pose and scene ambiguity can influence which solution the optimizer reaches.

Do not reuse cell size and optimization parameters across radically different physical scales without checking their meaning.

Do not regard every point as valid evidence. Outliers, moving objects, sparse cells, and limited overlap can distort the fit. Covariance estimates may need numerical treatment when local geometry is nearly degenerate.

Business use

NDT offers a useful registration option when large point clouds need to be aligned efficiently using a compact local statistical representation. Its suitability depends on sensor characteristics, geometric structure, pose initialization, and the precision needed after alignment.

Compare it with appropriate alternatives using representative scans and independent reference measurements. Measure translation and rotation error, failure rate, runtime, and sensitivity to overlap. For downstream dimensional decisions, ensure registration errors are not mistaken for real object changes or hidden by an overly forgiving fit criterion.

Check your understanding

A cell is much more spread horizontally than vertically. Should equal horizontal and vertical offsets necessarily receive the same distribution-based score?

Answer: No. Covariance accounts for directional spread, so the same ordinary distance can have different statistical compatibility.

Related terms

Covariance
Mahalanobis distance
Point-cloud registration
voxel grid
Gaussian distribution
rigid transformation
iterative closest point

Learn next

Plot an elongated two-dimensional Gaussian and compare ordinary distance with Mahalanobis distance. Then align a simple cloud with a known transform and vary cell size and initial pose while tracking recovery error.

Mahalanobis distance

Sources

Point Cloud Library, How to Use Normal Distributions Transform: https://pointclouds.org/documentation/tutorials/normal_distributions_transform.html . The official tutorial was read for its distribution representation, registration workflow, initialization, and scale-dependent parameters. The numerical cell example and corridor case are original illustrations.

Previous: Normal distribution Next: Normal equations