Iterative Closest Point (ICP) and other registration algorithms

Originally introduced in [BM92], the ICP algorithm aims at finding the transformation between a point cloud and some reference surface (or another point cloud), by minimizing the square errors between the corresponding entities. The ‘’iterative’’ of ICP comes from the fact that the correspondences are reconsidered as the solution comes closer to the error local minimum. As any gradient descent method, the ICP is applicable when we have a relatively good starting point in advance. Otherwise, it will be trapped into the first local minimum and the solution will be useless. In the field of mobile robots, ICP has been extensively employed to match 2D and 3D laser scans, a problem called ‘’scan matching’’. See also [Cen08, SHT09].

ICP algorithms in MRPT can take as input:

Two planar (2D) maps, either:
- A reference map as a cloud of points, and a map to be aligned as a cloud of points, or
- A reference map as an occupancy grid map, and a map to be aligned as a cloud of points.
Two 3D maps, both represented as clouds of points.

For the case of point maps, a KD-tree is used to accelerate the search of nearest neighbours. The ICP method is implemented in the class mrpt::slam::CICP. The output is a pdf (probability density function) of the relative pose between the maps, that is, an uncertainty bound is also computed associated to the optimal registration. An example of typical usage is (see also the example in the directory MRPT/samples/icp):

CICP icp;
// set ICP parameters:
icp.options.maxIterations = 50;
//...

// Reference map:
CSimplePointsMap refMap;

// Map to be aligned:
CSimplePointsMap alignMap;

// Initial guess, used in the first ICP iteration:
CPose2D initialGuess(0,0,0);
CPosePDFPtr pdf = icp.AlignPDF(
  &refMap, // Reference map
  &alignMap, // Map to be aligned
  initialGuess // Starting estimate
  );
CPose2D icpEstimateMean = pdf->getMeanVal();
cout << icpEstimateMean << endl;

The different ICP algorithms implemented in the MRPT C++ library (explained below) are:

The “classic ICP”.
A Levenberg-Marquardt iterative method.

The following animation shows how the threshold distance for establishing correspondences may have a great impact in the convergence (or not) of ICP:

Examples of usage

For 2D alignment, refer to the example: https://raw.github.com/MRPT/mrpt/master/samples/icp/
For `3D aligment <http://www.mrpt.org/3D-ICP_Example>`__ (full 6D poses), see the example: https://raw.github.com/MRPT/mrpt/master/samples/icp3D/

Provided algorithms

The “classic” ICP algorithm

This algorithm can be invoked in MRPT via the methods mrpt::slam::CICP::AlignPDF(), ::Align() (or their 3D equivalent versions) by setting ICP_algorithm = CICP::icpClassic in the structureCICP::options. The specific algorithm implemented in MRPT performs a kind of progressive refinement as it approaches convergence. If you want to disable the refinement stages, set the parameter ALFA=0. We show first a simplified version of the method in pseudo-code, then we’ll describe the more interesting parts and finally we give the complete list of parameters and their meaning:

i=0 // Iteration counter
P(i)=P0 // Initial guess given by the user
WHILE( i<max_iterations OR thres_dist>thres_dist_min )
  Matchings = ComputeMatching of m1 with m2 displaced by P(i) with thres_dist & thres_ang
  P(i+1) = LeastSquare(Matchings)
  IF (all components of |P(i+1)-P(i)|<1e-6)
    thres_dist *= alpha
    thres_ang *= alpha
    IF (thres_dist < thres_dist_min)
      BREAK; // End of the WHILE loop: we reached convergence
    END-IF
  END-IF
  i++;
END-WHILE

In words: the matching of the transformed point cloud with the reference point map is determined using thres_dist and thres_ang, then a solver is executed to obtain the 2D or 3D transformation that best matches those pairings. This is repeated until convergence and, if ALFA>0 (which is the default) the tresholds are reduced and the entire process repeated. The above algorithm is controlled by means of the following parameters in mrpt::slam::CICP::options:

TICPAlgorithm ICP_algorithm: …
bool onlyClosestCorrespondences: …
bool onlyUniqueRobust: …
unsigned int maxIterations: …
float thresholdDist,thresholdAng: When determining matchings between two point clouds, two nearby poins are considered as “candidate pairings” only if their distance is below ``thresholdDist + D*thresholdAng ``, which D being the distance of the point in the “to align” map to the map origin of coordinates. Mathematically, it models an uncertainty in the angular component of the pose between point clouds.
float ALFA: …
float smallestThresholdDist: …
float covariance_varPoints: …
bool skip_cov_calculation: …
bool doRANSAC: …
unsigned int ransac_minSetSize,ransac_maxSetSize,ransac_nSimulations: …
float ransac_mahalanobisDistanceThreshold: …
float normalizationStd: …
bool ransac_fuseByCorrsMatch: …
float ransac_fuseMaxDiffXY, ransac_fuseMaxDiffPhi: …
float kernel_rho: …
bool use_kernel: …
float Axy_aprox_derivatives: …
float LM_initial_lambda: …
uint32_t corresponding_points_decimation: Each point in m2 is tested for its nearest neighbor in m1 via a KD-tree. Queries to this KD-tree actually are the most time-consuming part of the entire ICP process. Thus is why it may be a good idea, when m2 is a dense point cloud, to downsample it. This parameter controls that downsampling (default=5), but can be changed to 1 to perform an exact matching search. However, the heuristics give very good results and the time improvement is drastic, so it’s recommended to set this parameter as high as possible while not degrading the accuracy of the result. Notice that only one out of “corresponding_points_decimation” points are matched against m1, but after each threshold scaling by “alfa”, the offset of these point index is shifted, so after a complete ICP alignment all points from “m2” have been considered. Only, that not all at the same time.

The Levenberg-Marquardt ICP algorithm

In this case, the only difference with the pseudo-code above is the replacement of this step:

LeastSquare(Matchings)

by:

NonLinearLeastSquare(Matchings)

where the optimizer that minimizes the average square error between pairings is implemented following the Levenberg-Marquardt algorithm. Jacobians are determined numerically to capture as well as possible the actual distribution of points. Credits for this algorithm are due to Dr. Paul Newman, on whose code was MRPT’s implementation based.

Optimizing sets of correspondences

Note

More general registration algorithms are provided in the more recent project MP2P_ICP

SE(2) Least Square Rigid transformation

Given a set of correspondences between two sets of points, this method computes the transformation that minimizes the square error. Implemented in tfest::leastSquareErrorRigidTransformation.

SE(3) Least Square Rigid transformation

Given a set of correspondences between two sets of points, this method computes the transformation that minimizes the square error. Implemented in tfest::leastSquareErrorRigidTransformation6D.

SE(2) Robust Rigid transformation

Given a set of correspondences between two sets of points, this method computes a Sum of Gaussians (SOG) over the potential transformations using a robust RANSAC stage. Implemented intfest::robustRigidTransformation.

Bibliography

BM92: Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, volume 1611, 586–606. International Society for Optics and Photonics, 1992.
Cen08: Andrea Censi. An icp variant using a point-to-line metric. In 2008 IEEE International Conference on Robotics and Automation, 19–25. Ieee, 2008.
SHT09: Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems, volume 2, 435. Seattle, WA, 2009.