Normal Distributions Transform

A point-cloud registration method that divides space into a grid, models each cell as a normal distribution, and aligns scans against it.

The Normal Distributions Transform (NDT) is a point-cloud registration method, first proposed by Biber and Straßer in 2003 for matching 2D laser scans and later extended to 3D. It divides a reference point cloud into a voxel grid, represents the points in each voxel as a normal distribution defined by a mean and covariance, and then optimizes the pose of the incoming point cloud so that its transformed points fall in the highest-probability regions of these distributions. Unlike Iterative Closest Point (ICP), which searches for point-to-point correspondences, NDT needs no nearest-neighbor search, which makes it more robust to a poor initial guess and to noise. It is widely used to localize a lidar sensor within a pre-built map, and is implemented in both the Point Cloud Library (PCL) and Autoware.

ExampleA self-driving car registers its current lidar scan against a pre-built point-cloud map using NDT to find its position within the map.

Also called

NDT, NDT Scan Matching, NDT Point Cloud Registration

Related

Point Cloud Registration · Iterative Closest Point · LiDAR SLAM · Relocalization · Point Cloud Library (PCL) · LiDAR

Sources

How to use Normal Distributions Transform - PCL Tutorials