Point cloud registration

Estimating a shared frame#

Two point clouds may describe the same surface from different sensor poses. Registration estimates the rigid-body transformation, or a more general deformation when the application permits one, that brings corresponding structure into alignment.

Registration can use known features, geometric descriptors, direct point or surface distances, or combinations of these. A coarse initial alignment is often followed by local refinement. The result can provide relative motion for simultaneous localization and mapping, align an object model to a depth scan, or combine several lidar views.

Iterative closest point is one family#

Iterative closest point repeatedly associates elements between two clouds, estimates a transformation that reduces an alignment error, and updates the moving cloud. The Point Cloud Library documentation demonstrates rigid alignment and reports a convergence flag, fitness score, and final transformation.

Variants use point-to-point, point-to-plane, or distribution-based errors. Generalized-ICP combines point-to-point and point-to-plane ideas in a probabilistic framework that models local surface structure in both scans. The paper reports better results than the tested standard and point-to-plane implementations; that comparison does not establish one variant as best for every sensor or scene.

Alignment can be plausible but wrong#

Local iterative methods depend on their initial pose and can converge to a wrong local minimum. Repetitive geometry, few overlapping points, moving objects, sparse measurements, and featureless surfaces can make the transformation ambiguous. A low residual may still describe the wrong alignment if the correspondences are wrong.

Filters, outlier rejection, uncertainty estimates, and a separate place-recognition or feature-matching stage can reduce these risks. Registration accuracy also depends on sensor calibration and timing. It should not be interpreted as a complete localization system or as proof that a scene is static.