Introduction Scan registration Experimental environment Experiments and results Conclusions
3D Mapping the Kvarntorp Mine— A Field Experiment for Evaluation of 3D
Scan Matching Algorithms
Martin Magnusson1, Andreas Nüchter2,Christopher Lörken2, Achim J. Lilienthal1, and
Joachim Hertzberg2
1AASS, Örebro University2KBS, Osnabrück University
September 26, 2008
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Outline
1 IntroductionOutlineBenchmarking
2 Scan registrationICPNDTExperimental setup
3 Experimental environmentThe Kvarntorp mine
4 Experiments and resultsValley of convergenceOutlier countCumulative errorExecution time
5 Conclusions
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Benchmarking
Unbiased comparisons
We are all biased.
Novel algorithms are often compared to previous workwithout the experience and thoroughness of the originalauthors.
We have attempted an unbiased comparison.
Two groups ran their implementations using their hands-onknowledge about parameter selection, using the samehardware and the same input data.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Benchmarking
Repeatable and evaluatable comparisons
Repeatable test:publicly available data,specify all relevant parameters,show some cases where the method doesn’t perform well.
Performance metrics:quantitative error measures,execution time.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
What is scan registration?
The process of aligning overlapping scans (point clouds).
First step of building a metric map.
Find the correct pose of the current scan by comparingsurface shapes with the previous scan.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
ICP
ICP: the iterative corresponding point algorithm
Find closest point in previous scan for each point in currentscan.
Minimise point-to-point distances.
Iterate until convergence.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
ICP
ICP specifics
Approximate kd trees
Cached kd treesAndreas Nüchter, Kai Lingemann, Joachim Hertzberg, andHartmut Surmann, 6D SLAM — 3D mapping outdoorenvironments, J. Field Robotics 24 (2007), no. 8–9,699–722.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
ICP
ICP specifics
Approximate kd trees
Cached kd treesAndreas Nüchter, Kai Lingemann, Joachim Hertzberg, andHartmut Surmann, 6D SLAM — 3D mapping outdoorenvironments, J. Field Robotics 24 (2007), no. 8–9,699–722.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
NDT
NDT: the normal distributions transform
Probabilistic approach to scanregistration.
Represent the reference surfacewith probability density functions(normal distributions).
Registration = Finding thefunction optimum whenevaluating the PDFs at the pointsin the current scan = Just anordinary numerical optimisationproblem.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
NDT
NDT variants
Discretising space into aregular grid gives a fewproblems:
finding the best cell size,limited field of influence,discontinuities at cellboundaries.
Solutions:
iterative NDT,linked cells,trilinear interpolation.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
NDT
NDT variants
Discretising space into aregular grid gives a fewproblems:
finding the best cell size,limited field of influence,discontinuities at cellboundaries.
Solutions:iterative NDT,
linked cells,trilinear interpolation.
Martin Magnusson, Achim J. Lilienthal, and Tom Duckett,Scan registration for autonomous mining vehicles using3D-NDT, J. Field Robotics 24 (2007), no. 10, 803–827.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
NDT
NDT variants
Discretising space into aregular grid gives a fewproblems:
finding the best cell size,limited field of influence,discontinuities at cellboundaries.
Solutions:iterative NDT,
linked cells,trilinear interpolation.
Martin Magnusson, Achim J. Lilienthal, and Tom Duckett,Scan registration for autonomous mining vehicles using3D-NDT, J. Field Robotics 24 (2007), no. 10, 803–827.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
NDT
NDT variants
Discretising space into aregular grid gives a fewproblems:
finding the best cell size,limited field of influence,discontinuities at cellboundaries.
Solutions:iterative NDT,
linked cells,trilinear interpolation.
Martin Magnusson, Achim J. Lilienthal, and Tom Duckett,Scan registration for autonomous mining vehicles using3D-NDT, J. Field Robotics 24 (2007), no. 10, 803–827.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
NDT
NDT variants
Discretising space into aregular grid gives a fewproblems:
finding the best cell size,limited field of influence,discontinuities at cellboundaries.
Solutions:iterative NDT,linked cells,
trilinear interpolation.
Martin Magnusson, Achim J. Lilienthal, and Tom Duckett,Scan registration for autonomous mining vehicles using3D-NDT, J. Field Robotics 24 (2007), no. 10, 803–827.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
NDT
NDT variants
Discretising space into aregular grid gives a fewproblems:
finding the best cell size,limited field of influence,discontinuities at cellboundaries.
Solutions:iterative NDT,linked cells,trilinear interpolation.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
NDT
NDT variants
Discretising space into aregular grid gives a fewproblems:
finding the best cell size,limited field of influence,discontinuities at cellboundaries.
Solutions:iterative NDT,linked cells,trilinear interpolation.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Experimental setup
Parameters
NDTIterative discretisation: 2 m, 1 m, 0.5 m.Linked cells.Newton optimisation: max step size |∆~p| = 0.2, convergencethreshold |∆~p| < 10−6.
ICPCached kd trees, 10 points per bucket.Distance threshold 0.5 m.Convergence treshold |∆~p| < 10−6
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
The Kvarntorp mine
Location
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
The Kvarntorp mine
Location
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
The Kvarntorp mine
Location
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
The Kvarntorp mine
Location
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
The Kvarntorp mine
Location
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
The Kvarntorp mine
Location
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
The Kvarntorp mine
Location
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
The Kvarntorp mine
Characteristics
Limestone mine (no longer in production).
Wide tunnels and relatively planar walls and ceiling.
Data collected in a part of the mine used for testing ofautomated loading vehicles.
Smooth artificial walls erected along some tunnels to makethem narrower and more challenging for tunnel followingbehaviours.
Tunnels are about 5 m high and 5 to 10 m wide.
Scans subsampled to 8000 points, around 20 m long.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Valley of convergence
Surveying the valley of convergence
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Valley of convergence
Plot legend
Use 7 × 7 translation offsets from-2 m to +2 m.
For each translation offset, use 9rotation offset from -80◦to +80◦.
Success: position within 1.0 m or0.2 m and orientation within 5◦
of reference.
0o
180o
90o
-90o
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Valley of convergence
NDT vs. ICP (loose threshold)
-2
-1
0
1
2
-2 -1 0 1 2
-2
-1
0
1
2
-2 -1 0 1 2
Success rate: 77% for NDT, 30% for ICP.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Valley of convergence
NDT vs. ICP (strict threshold)
-2
-1
0
1
2
-2 -1 0 1 2
-2
-1
0
1
2
-2 -1 0 1 2
Success rate: 37% for NDT, 13% for ICP.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Valley of convergence
NDT with and without trilinear interpolation(loose threshold)
-2
-1
0
1
2
-2 -1 0 1 2
-2
-1
0
1
2
-2 -1 0 1 2
Success rate: 77% for NDT, 95% with trilinear interpolation.M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Valley of convergence
NDT with and without trilinear interpolation(strict threshold)
-2
-1
0
1
2
-2 -1 0 1 2
-2
-1
0
1
2
-2 -1 0 1 2
Success rate: 37% for NDT, 95% with trilinear interpolation.M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Outlier count
Outlier count
Count how many times the initial pose estimate (fromodometry) had to be manually altered for convergence.Results:
ICP: 1 (scan nr 33)NDT: 1 (scan nr 23)Trilinear NDT : none
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Cumulative error
Cumulative error at loop closure
Measure accumulated errorafter registering 55 scans(160 m trajectory).
Reference: register last scanof loop to first scan.
(Only a rough estimate ofregistration performance.)
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Cumulative error
Trilinear NDT
Accumulated translation error: 3.99 m
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Cumulative error
ICP
Accumulated translation error: 2.97 m
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Cumulative error
NDT
Accumulated translation error: 2.26 m
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Execution time
Execution time
0
5
10
15
20
25T
ime
(s)
ICP 0
5
10
15
20
25
Tim
e (s
)
NDT 0
5
10
15
20
25
Tim
e (s
)
Trilinear NDT
All three algorithms can be performed in “real time”. Vehiclemotion and (even more so) scanning speed are the mainbottlenecks.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Summary of results
ICP NDT Trilinear NDTConvergence (1.0 m) 30% 77% 95%Convergence (0.2 m) 13% 37% 95%Outliers 1/55 1/55 0/55Time 4.8 s 2.4 s 9.5 s
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Summary of results
ICP NDT Trilinear NDTConvergence (1.0 m) 30% 77% 95%Convergence (0.2 m) 13% 37% 95%Outliers 1/55 1/55 0/55Time 4.8 s 2.4 s 9.5 s
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Summary of results
ICP NDT Trilinear NDTConvergence (1.0 m) 30% 77% 95%Convergence (0.2 m) 13% 37% 95%Outliers 1/55 1/55 0/55Time 4.8 s 2.4 s 9.5 s
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Summary of results
ICP NDT Trilinear NDTConvergence (1.0 m) 30% 77% 95%Convergence (0.2 m) 13% 37% 95%Outliers 1/55 1/55 0/55Time 4.8 s 2.4 s 9.5 s
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
Conclusions
Trilinear NDT converges from the largest range of poses.
ICP has stable but narrow valley of convergence.
Standard NDT is the fastest.
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms
Introduction Scan registration Experimental environment Experiments and results Conclusions
x
M. Magnusson, A. Nüchter, C. Lörken, A. J. Lilienthal, and J. Hertzberg
3D Mapping the Kvarntorp Mine — A Field Experiment for Evaluation of 3D Scan Matching Algorithms