Point Cloud Registration

Finding the rotation and translation that lines up two point clouds within one shared coordinate frame.

Point cloud registration solves for a spatial transform that aligns a source point cloud onto a target one; rigid registration allows only rotation and translation, while non-rigid registration also permits deformation. It's needed whenever a scene is scanned from different viewpoints or times and the results need to be stitched into a full model, when estimating how far a robot has moved in lidar odometry (often called scan matching in that context), or when aligning an object's model onto an observed point cloud to recover its pose. A common pipeline does coarse global registration first, using local features such as FPFH to find correspondences plus RANSAC to reject bad matches, and then fine registration with iterative closest point (ICP), which repeatedly finds the nearest points and solves for the optimal transform — though it depends on a reasonably good initial guess or it gets stuck in a local optimum. Open3D's ICP implementation also reports an overlap ratio (fitness) and inlier RMSE to judge the result's quality.

ExampleA wrist camera photographs the same cup from two angles. FPFH plus RANSAC first coarsely aligns the two point clouds, then ICP refines the alignment to stitch together a more complete shape of the cup.

Also called

Point Cloud Alignment, Scan Matching, Point Set Registration

Related

Iterative Closest Point · Fast Point Feature Histograms · Random Sample Consensus · Normal Distributions Transform · 6D Object Pose Estimation · Simultaneous Localization and Mapping

Sources

Wikipedia: Point-set registration
Open3D: ICP registration