Introduction
The Normal Distributions Transform (NDT) is a technique which transforms a raw point cloud into a grid-based representation, where each grid cell is approximated with a multivariate Gaussian distribution.
The purpose of this representation is to provide a compact form for representing the point cloud. Doing so also permits the holistic representation of the scan matching, or scan registration problem as a differentiable optimization problem.
This permits us to use the NDT algorithm to do point cloud-based localization.
Literature Review
The following is a list of papers relating specifically to the NDT technique for point cloud scan matching:
- Scan Registration for Autonomous Mining Vehicles Using 3D-NDT
- Extends the original NDT paper from 3 DoF scan registration to a full 6 DoF registration
- This can be thought of as Point-to-Distribution NDT (P2D)
- The Three-Dimensional Normal-Distributions Transform – an Efficient Representation for Registration, Surface Analysis, and Loop Detection
- Thesis for NDT, covers 2D-NDT, 3D-NDT, Color-NDT
- Also covers variations on NDT representation:
- Trilinear interpolation
- Linked nodes, etc.
- Has a variety of empirical results:
- NDT is faster and more consistent than ICP
- NDT is somewhat more robust than ICP, but still needs good initializations
- Color-NDT is better than plain NDT
- Also covers applications, e.g. loop-closure
- Fast and accurate scan registration through minimization of the distance between compact 3D NDT representations
- Introduces Distribution-to-Distribution NDT (D2D)
- Faster than P2D-NDT
- Beyond Points: Evaluating Recent 3D Scan-Matching Algorithms
- Compares multiple NDT techniques, and various tweaks on the algorithm, i.e.:
- (Trilinear) interpolation for handling boundary effects
- Iterative subdivision to improve robustness
- Linked nodes for handling outliers/missing data
- P2D is more robust for unstructured environments and small overlaps between scans compared to D2D
- D2D is faster and less sensitive to poor initial estimates
- Point cloud registration from local feature correspondences—Evaluation on challenging datasets
- Compares ICP variants and NDT
- NDT is good at handling translation, and among the fastest of techniques
- NDT is so-so at handling rotations
- Evaluation of 3D Registration Reliability and Speed – A Comparison of ICP and NDT
- Compares ICP with NDT variants
- Introduces trilinear interpolation for NDT - results in more robustness
- NDT is slightly faster than ICP, generally more robust
- Incorporating Ego-motion Uncertainty Estimates in Range Data Registration
- Adds a motion constraint to the optimization objective
- Slower than normal D2D
- However, much more performant on low-information environments (i.e. endless hallway)
- Semantic-assisted 3D Normal Distributions Transform for scan registration in environments with limited structure
- Generate several NDTs based on metric/feature-based partitions (which also discard outliers), extend optimization objective to handle these multiple NDTs
- Shows strong speed and accuracy results