pcl ::registration ::registration initial alignment of point clouds reﬁning initial alignments ......

Registration Iterative Closest Points (ICP) Initial Alignment Example/Tutorial Code

PCL :: RegistrationInitial alignment of point clouds

Refining initial alignmentsSeptember 25, 2011


Outline

1. Registration

2. Registration using the Iterative Closest Points (ICP)

3. Feature-Based Initial Alignment

4. Example/Tutorial Code

Dirk Holz / PCL :: Registration


Registration

Wanted: transformation that aligns one point cloud to another.

I for initially aligning point clouds (based on features)I for refining initial alignments

(using the Iterative Closest Point (ICP) Algorithm)Dirk Holz / PCL :: Registration


ICP (1/6)

Registration using the Iterative Closest Point (ICP) Algorithm

Given an input point cloud and a target point cloud1. determine pairs of corresponding points,2. estimate a transformation that minimizes the distances

between the correspondences,3. apply the transformation to align input and target.



ICP (2/6)

I Given:Two n-dimensional sets of points – the model setM = {mi |mi ∈ Rn, i = 1, . . . ,Nm} and the data setD = {dj |dj ∈ Rn, j = 1, . . .Nd}

I Wanted:A rotation R and a translation ∆t that map D onto M.

I handled as an optimization problem =⇒ Minimizingmapping error E

E (R,∆t) =Nm∑i=1

Nd∑j=1

wi,j∣∣∣∣mi −

(Rdj + ∆t

)∣∣∣∣2 (1)

I Weighting factor wi,j encodes point correspondences,wi,j = 1 for correspondence(mi ,di), 0 otherwise

I Correspondence from Neighbor SearchDirk Holz / PCL :: Registration


ICP (3/6)

The ICP APIpcl::IterativeClosestPoint<InType, OutType> icp;

Provide a pointer to the input point cloudicp.setInputCloud (input_cloud);

Provide a pointer to the target point cloudicp.setInputTarget (target_cloud);

Align input to target to obtainicp.align (aligned_cloud);

Eigen::Matrix4f transformation = icp.

getFinalTransformation ();

I the aligned cloud (transformed copy of input cloud),I and transformation used for alignment.



ICP (4/6)

The ICP API: Termination CriteriaI Max. number of iteration steps→ set via setMaximumIterations(nr_iterations)

I Convergence: Estimated transformation doesn’t change(the sum of differences between current and lasttransformation is smaller than a user-defined threshold)→ set via setTransformationEpsilon(epsilon)

I A solution was found (the sum of squared errors is smallerthan a user-defined threshold)→ set via setEuclideanFitnessEpsilon(distance)



ICP (5/6)

The ICP API: Problems with ICP

False correspondences negatively affect the alignment(the algorithms gets caught in local minima).

I Maximum distance between correspondences:icp.setMaxCorrespondenceDistance (distance);

I Use RANSAC to neglect false correspondencesicp.setRANSACOutlierRejectionThreshold (distance);

I Model: Transformation (estimated on 3 samples)I Inliers: Points whose distance to the corresponding is

below the given threshold



ICP (6/6)

The ICP API: Problems with ICP

The ICP algorithms needs a rough initial alignment.I Use Features to get an initial alignment!

→ pcl::SampleConsensusInitialAlignment



Initial Alignment (1/7)

1. Compute sets of keypoints2. Compute (local) feature descriptors (e.g. FPFH)3. Use SAC-based approach to find initial alignment

3.1 Take 3 random correspondence pairs3.2 Compute transformation for these pairs3.3 Apply transformation to all source points, and determine

inliers4. Use best transformation for initial alignment, and ICP for

refinement

*Dirk Holz / PCL :: Registration



points on similar surfaces

*

0

5

10

15

20

25

30

35

0 2 4 6 8 10 12 14 16

Rat

io o

f poi

nts

in o

ne b

in (

%)

Bins

Persistent Feature Points Histograms

P1

Q1

0

5

10

15

20

25

30

35

0 2 4 6 8 10 12 14 16

Rat

io o

f poi

nts

in o

ne b

in (

%)

Bins


P2

Q2

0

5

10

15

20

25

30

35

0 2 4 6 8 10 12 14 16

Rat

io o

f poi

nts

in o

ne b

in (

%)

Bins


P3

Q3




*



Initial Alignment (4/7)Outdoor Example: Non-Linear Optimization

*Dirk Holz / PCL :: Registration



The SampleConsensusInitialAlignment APIpcl::SampleConsensusInitialAlignment<PointT, PointT,

DescriptorT> sac;

Provide a pointer to the input point cloud and featuressac.setInputCloud (source_points);

sac.setSourceFeatures (source_descriptors);

Provide a pointer to the target point cloud and featuressac.setInputTarget (target_points);

sac.setTargetFeatures (target_descriptors);

Align input to target to obtainsac.align (aligned_cloud);

Eigen::Matrix4f transformation = sac.

getFinalTransformation ();



Example/Tutorial Code (1/5)

Initial alignment of point clouds

1 pcl::SampleConsensusInitialAlignment<PointT, PointT,DescriptorT> sac;

2 sac.setMinSampleDistance (min_sample_distance);3 sac.setMaxCorrespondenceDistance (

max_correspondence_distance);4 sac.setMaximumIterations (nr_iterations);5

6 sac.setInputCloud (source_points);7 sac.setSourceFeatures (source_descriptors);8

9 sac.setInputTarget (target_points);10 sac.setTargetFeatures (target_descriptors);11

12 PointCloudPtr aligned_source(new PointCloud);13 sac.align (*aligned_source);14 Eigen::Matrix4f initial_T = sac.getFinalTransformation

());



Example Code (2/5)

Refining initial alignments using ICP

1 pcl::IterativeClosestPoint<PointT, PointT> icp;2 icp.setMaxCorrespondenceDistance (distance);3 icp.setRANSACOutlierRejectionThreshold (distance);4 icp.setTransformationEpsilon (transformation_epsilon);5 icp.setMaximumIterations (max_iterations);6

7 icp.setInputCloud (aligned_source); // from (1)8 icp.setInputTarget (target_points);9

10 PointCloud registration_output;11 icp.align (registration_output);12

13 Eigen::Matrix4f refined_T =14 icp.getFinalTransformation () * initial_T);



Example Code (3/5)

Example: Correspondences in initial alignment$ ./correspondence_viewer ../../data/robot/robot0

../../data/robot/robot1 -n 5



Example Code (4/5)

Example: Initial alignment + refinement$ ./test_registration ../../data/robot/robot0 ../../

data/robot/robot1 -i 0.025,0.01,500 -r 0.05,0.05,0,100



Example Code (5/5)

Example: Final model = robot0 + robot1$ ./test_registration ... -s robot_registered.pcd

$ pcd_viewer robot_registered.pcd


pcl ::registration ::registration initial alignment of point clouds reﬁning initial alignments ......

Documents