simpleICP

This repo contains implementations of a rather simple version of the Iterative Closest Point (ICP) algorithm in various languages.

Currently, an implementation is available for:

Language	Code	Main dependencies
C++	Link	nanoflann, Eigen, cxxopts
Julia	Link	NearestNeighbors.jl
Matlab	Link	Statistics and Machine Learning Toolbox
Octave	Link
Python	Link	NumPy, SciPy, lmfit, pandas

I've tried to optimize the readability of the code, i.e. the code structure is as simple as possible and tests are rather rare.

The C++ version can be used through a cli interface.

Also available at:

Matlab:
Python:

Features of the ICP algorithm

Basic features

The following basic features are implemented in all languages:

Usage of the signed point-to-plane distance (instead of the point-to-point distance) as error metric. Main reasons:
- higher convergence speed, see e.g. here and here
- better final point cloud alignment (under the assumption that both point clouds are differently sampled, i.e. no real point-to-point correspondences exist)
Estimation of a rigid-body transformation (rotation + translation) for the movable point cloud. The final transformation is given as homogeneous transformation matrix H:
```
H = [R(0,0) R(0,1) R(0,2)   tx]
    [R(1,0) R(1,1) R(1,2)   ty]
    [R(2,0) R(2,1) R(2,2)   tz]
    [     0      0      0    1]
```
where R is the rotation matrix and tx, ty, and tz are the components of the translation vector. Using H, the movable point cloud can be transformed with:
```
Xt = H*X
```
where X is a 4-by-n matrix holding in each column the homogeneous coordinates x, y, z, 1 of a single point, and Xt is the resulting 4-by-n matrix with the transformed points.
Selection of a fixed number of correspondences between the fixed and the movable point cloud. Default is correspondences = 1000.
Automatic rejection of potentially wrong correspondences on the basis of
1. the median of absolute deviations. A correspondence i is rejected if |dist_i-median(dists)| > 3*sig_mad, where sig_mad = 1.4826*mad(dists).
2. the planarity of the plane used to estimate the normal vector (see below). The planarity is defined as P = (ev2-ev3)/ev1 (ev1 >= ev2 >= ev3), where ev are the eigenvalues of the covariance matrix of the points used to estimate the normal vector. A correspondence i is rejected if P_i < min_planarity. Default is min_planarity = 0.3.
After each iteration a convergence criteria is tested: if the mean and the standard deviation of the point-to-plane distances do not change more than min_change percent, the iteration is stopped. Default is min_change = 1.
The normal vector of the plane (needed to compute the point-to-plane distance) is estimated from the fixed point cloud using a fixed number of neighbors. Default is neighbors = 10.
The point clouds must not fully overlap, i.e. a partial overlap of the point cloud is allowed. An example for such a case is the Bunny dataset, see here. The initial overlapping area between two point clouds can be defined by the parameter max_overlap_distance. More specifically, the correspondences are only selected across points of the fixed point cloud for which the initial distance to the nearest neighbor of the movable point cloud is <= max_overlap_distance.

Extended features

The extended features are currently not implemented in all languages. The differences are documented in the following table:

Feature	C++	Julia	Matlab	Octave	Python
observation of rigid-body transformation parameters	no	no	no	no	yes

Extended feature: observation of rigid-body transformation parameters

This is useful in at least these cases:

If only a subset of the 6 rigid-body transformation parameters should be estimated. This can be accomplished by setting the weight of individual parameters to infinite, see example below.
If all or a subset of the 6 rigid-body transformation parameters have been directly observed in any other way, e.g. by means of a manual measurement.
If estimates for the rigid-body transformation parameters exist, e.g. from a previous run of simpleICP. In this case the observation weight should be set (according to the theory of least squares adjustments) to w = 1/observation_error^2 whereby the observation_error is defined as std(observation_value). The observation error of all parameters is reported by simpleICP as "est.uncertainty" in the logging output.

This feature introduces two new parameters: rbp_observed_values and rbp_observation_weights. Both parameters have exactly 6 elements which correspond to the rigid-body transformation parameters in the following order:

alpha1: rotation angle around the x-axis
alpha2: rotation angle around the y-axis
alpha3: rotation angle around the z-axis
tx: x component of translation vector
ty: y component of translation vector
tz: z component of translation vector

The rigid-body transformation is defined in non-homogeneous coordinates as follows:

Xt = RX + t

where X and Xt are n-by-3 matrices of the original and transformed movable point cloud, resp., t is the translation vector, and R the rotation matrix. R is thereby defined as:

R = [ca2*ca3               -ca2*sa3                sa2    ]
    [ca1*sa3+sa1*sa2*ca3    ca1*ca3-sa1*sa2*sa3   -sa1*ca2]
    [sa1*sa3-ca1*sa2*ca3    sa1*ca3+ca1*sa2*sa3    ca1*ca2]

with the substitutions:

sa1 := sin(alpha1), ca1 := cos(alpha1)
sa2 := sin(alpha2), ca2 := cos(alpha2)
sa3 := sin(alpha3), ca3 := cos(alpha3)

The two parameters rbp_observed_values and rbp_observation_weights can be used to introduce an additional observation to the least squares optimization for each transformation parameter:

residual = observation_weight * (estimated_value - observed_value)

Example which demonstrates the most important combinations:

# parameters:              alpha1   alpha2   alpha3   tx      ty     tz
rbp_observed_values =     (10.0     0.0     -5.0      0.20   -0.15   0.0)
rbp_observation_weights = (100.0    0.0      0.0      40.0    40.0   inf)

Consequently:

alpha1: is observed to be 10 degrees with an observation weight of 100.
alpha2: is not observed since the corresponding weight is zero. However, the observed value is used as initial value for alpha2 in the non-linear least squares optimization.
alpha3: is also not observed, but has an initial value of -5 degrees.
tx: is observed to be 0.20 with an observation weight of 40.
ty: is observed to be -0.15 with an observation weight of 40.
tz: is observed to be 0 with an infinite observation weight, i.e. this parameter becomes a constant and is fixed to be exactly the observation value. Thus, in this case only 5 (out of 6) rigid-body transformation parameters are estimated.

Output

All implementations generate the same screen output. This is an example from the C++ version for the Bunny dataset:

$ run_simpleicp.sh
Processing dataset "Dragon"
Create point cloud objects ...
Select points for correspondences in fixed point cloud ...
Estimate normals of selected points ...
Start iterations ...
Iteration | correspondences | mean(residuals) |  std(residuals)
   orig:0 |             767 |          0.0001 |          0.3203
        1 |             767 |         -0.0061 |          0.2531
        2 |             773 |         -0.0035 |          0.1669
        3 |             771 |         -0.0008 |          0.0835
        4 |             741 |         -0.0006 |          0.0196
        5 |             762 |          0.0000 |          0.0025
        6 |             775 |          0.0001 |          0.0022
Convergence criteria fulfilled -> stop iteration!
Estimated transformation matrix H:
[    0.998696     0.052621    -0.034179    -0.206737]
[   -0.052090     0.999028     0.020119    -0.408088]
[    0.034822    -0.018663     0.999436    -0.593361]
[    0.000000     0.000000     0.000000     1.000000]
Finished in 1.729 seconds!

Test data sets

The test data sets are included in the data subfolder. An example call for each language can be found in the run_simpleicp.* files, e.g. run_simpleicp.jl for the julia version.

Dataset	pc1 (no_pts)	pc2 (no_pts)	Overlap	Source
Dragon	pc1 (100k)	pc2 (100k)	full overlap	The Stanford 3D Scanning Repository
Airborne Lidar	pc1 (1340k)	pc2 (1340k)	full overlap	Airborne Lidar flight campaign over Austrian Alps
Terrestrial Lidar	pc1 (1250k)	pc2 (1250k)	full overlap	Terrestrial Lidar point clouds of a stone block
Bunny	pc1 (21k)	pc2 (22k)	partial overlap	The Stanford 3D Scanning Repository

Benchmark

These are the runtimes on my PC for the data sets above:

Dataset	C++	Julia	Matlab	Octave*	Python
Dragon	0.16s	3.99s	1.34s	95.7s	4.51s
Airborne Lidar	3.98s	5.38s	15.08s	-	16.49s
Terrestrial Lidar	3.62s	5.22s	13.24s	-	14.45s
Bunny	0.13s	0.38s	0.37s	72.8s	4.20s

For all versions the same input parameters (correspondences, neighbors, ...) are used.

* Unfortunately, I haven't found an implementation of a kd tree in Octave (it is not yet implemented in the Statistics package). Thus, a (very time-consuming!) exhaustive nearest neighbor search is used instead. For larger datasets the Octave timings are missing, as the distance matrix does not fit into memory.

References

Please cite related papers if you use this code:

@article{glira2015a,
  title={A Correspondence Framework for ALS Strip Adjustments based on Variants of the ICP Algorithm},
  author={Glira, Philipp and Pfeifer, Norbert and Briese, Christian and Ressl, Camillo},
  journal={Photogrammetrie-Fernerkundung-Geoinformation},
  volume={2015},
  number={4},
  pages={275--289},
  year={2015},
  publisher={E. Schweizerbart'sche Verlagsbuchhandlung}
}

Related projects

globalICP: A multi-scan ICP implementation for Matlab

rstallman / simpleICP