www.data61.csiro.au

REFERENCES

[Bosse and Zlot, 2013]M. Bosse and R. Zlot. Place recognition using keypointvoting in large 3D lidar datasets. In 2013 IEEE International Conference onRobotics and Automation, pages 2677–2684, May 2013.

ACKNOWLEDGEMENTS

The authors gratefully acknowledge funding of the projectbythe CSIRO and QUT. The institutional support of CSIROandQUT, and the help of several individual members of staffinthe Autonomous Systems Laboratory including TomLowe,Gavin Catt and Mark Cox are greatly appreciated.

Probabilistic Surfel Fusion for Dense LiDAR Mapping

Problem Statement

We present a new approach for dense LiDAR mapping using probabilistic surfel fusion. The proposed system is capableof reconstructing a large-scale high-quality dense surface element (surfel) map from spatially redundant multiple views.

Chanoh Park1,2, Soohwan Kim1, Peyman Moghadam1,2, Clinton Fookes2, Sridha Sridharan2

1Autonomous Systems Laboratory, CSIRO Data61, Brisbane, Australia2Queensland University of Technology, Brisbane, Australia

Surfel Matching

Experiment Results

FOR FURTHER INFORMATION

Chanoh Park

e Chanoh.Park@data61.csiro.au

Peyman Moghadam, Soohwan Kim

e {Peyman.Moghadam, Soohwan.Kim}@data61.csiro.au

Proposed Method

Figure 1: Degenerate surfel normal caused by a degenerate points shape.

Local Mapping Global Mapping

Local SLAM

Module

Dense Surfel Fusion

Localization

and Surfel Integration

Sparse Surfel

Map

Dense Surfel

MapDense Surfels

Radius Search

Map Update

Active Area

Map Update

Multi-Resolution

Sparse Surfels

LiDAR

Transformation

Raw Points

Cloud

3D Ellipsoidal Surfel Map 2D Disk Surfel Mapfrom Multi-resolutional Voxel Hassing from Nearest Neighbor Searching

LiDAR

Beam

Surfel

Surfel normal

Distance

Incident angle

3D position uncertainty

Possible distribution of normal vector

Tangentiality reinforcement

Before After

Unit sphere surface

Normal

Uncertainty

Black line : fused normal uncertainty on the unit sphere

Surf

elPo

siti

on

U

nce

rtai

nty

Mo

del

ing

LiDAR

Beam

Surfel

Surfel normal

Surfel normal uncertainty

Surf

elN

orm

al

Un

cert

ain

ty M

od

elin

g

Conclusion

Figure 6: Illustration of surfel matchingproblem between a local map surfel and theglobal map surfels.

Figure 7: Proposed two staged matching algorithm. The step1controls map resolution whereas the step2 reduces map noise bysearching deeper along the LiDAR beam direction.

realstructure

Uncertainty along the normal direction

𝜎2

realstructure

r1 r2r3 r4Global Map Surfels

Global map

Local map

Deeper searchingwith a Mahalanobisdistance threshold

Higher Resolutionwith a Euclidian

distance threshold

Step1: Resolutional Matching

Step2: Depth Matching

Matching candidates

Side view of the matching candidates

The proposed two staged matching algorithm accurately finds the matched surfel of the new input surfel from the local map in the global map.

Probabilistic dense surfel fusion for LiDAR is proposed. Our method showeddenser but lesser noise level in building a dense surfel map. Also, theproposed method has an advantage on long-term SLAM applications.

Figure 9: The experimental handheld3D spinning LiDAR for mobile mapping.

Figure 8: Surfel statistics and uncertainties. [left] The number of surfels and theaverage number of fusion per surfel, [right] Uncertainties of surfel positions andnormal vectors

When building the dense surfel map fromLiDAR points cloud, there are two mainissues. The first issue is surfel degeneracy in anormal direction of a surfel which causesincorrect normal directions. The other issue isthat surfel matching is less accurate or notstraightforward in the traditional methods.Radius search cannot handle sensor noiseefficiently and it is difficult to control thesurface resolution in the uncertainty basedmethod.

System Overview

Map Representations

Sensor Noise Modelling and Surfel Uncertainty

Figure 4: (a) Example of a 3D ellipsoid surfel map with a 60cm resolution and (b) a 2D disk surfel map with a 1cm resolution.Both are color-coded by normal directions. Recognize the ceiling and the floor in blue, and objects and walls in orange andgreen.

3D ellipsoid surfel mapis faster and morerobust to run lo-calization, and dense2D disk surfel map isdenser and more ac-curate to 3D recon-struct the environ-ment.

Figure 2: Surfel matching problem in aradius search method(left) and uncertaintybased method(right)

Figure 3: The proposed system is composed of two main stages. Local mapping stage processes the raw LiDAR data andcreates local maps. The global mapping stage build a globally consistent map by merging them.

All matched surfels aremerged and updated by aBayesian formula.

Figure 11: (a) Raw camera image of the office in thered circle. (b) Synthesized image from the surfelmap. (c) Surfels map colored with normal direction.(d) Depth image.

Figure 10: Reconstructed 3D surfel map of an 20x20 meter office witha color fusion by camera images.

(a) (b)

(c) (d)

Figure 5: The Illustration of the relationshipbetween the underling points cloud of a surfel andthe normal uncertainty

For surfel normal update,an additional step for tan-gentiality reinforcement isrequired.

Surfel Fusion