Fast Plane Extraction in Organized Point Clouds Using … · 2019. 10. 22. · Fast Plane Extraction in Organized Point Clouds Using Agglomerative Hierarchical Clustering Chen Feng1),

Fast Plane Extraction in Organized Point Clouds Using Agglomerative Hierarchical Clustering

Chen Feng1), Yuichi Taguchi2), and Vineet R. Kamat1)

1) University of Michigan, USA

2) Mitsubishi Electric Research Labs, USA

June 4, 2014

Outline

Introduction

Overview

Breakdown

Experiments

Conclusions

2

Introduction Overview Breakdown Experiments Conclusions

• Real-time plane extraction is crucial to various applications in robotics, computer vision, and 3D modeling:

– Table-top object manipulation

– Landmarks for SLAM

– Compact and semantic scene modeling

3


• Real-time plane extraction is crucial to various applications in robotics, computer vision, and 3D modeling:

– Table-top object manipulation

– Landmarks for SLAM

– Compact and semantic scene modeling

• We present an efficient and reliable fast plane extraction algorithm applicable to organized point clouds, such as depth maps obtained by Kinect sensors.

3


• Previous Work

4


• Previous Work – RANSAC-based

• “Surfels” from Hough Transform (Oehler et al. 2011)

• RANSAC on local region (Taguchi et al. 2013; Hulik et al. 2012; Lee et al. 2012)

4





– Region-grow-based • Point-plane distance/MSE threshold (Hahnel et al. 2003; Poppinga et

al. 2008)

• Surface normal deviation threshold (Holz & Behnke 2012)

• Line segments grow (Georgiev et al. 2011)

4






al. 2008)



– Graph-based (Strom et al. 2010; Wang et al. 2013)

4






al. 2008)



– Graph-based (Strom et al. 2010; Wang et al. 2013)

– Other • Normal space clustering (Holz et al. 2011)

• Gradient-of-depth feature (Enjarini et al. 2012)

4


• Previous Work

5

0

5

10

15

20

25

30

35

40Average FPS for VGA (640x480) point clouds


• Analogy to Line Regression (Nguyen et al. 2005; April Robotics

Toolkit, 2010)

– Exploit the neighborhood information given by the order of points

6



Toolkit, 2010)


6

2D point sequences



Toolkit, 2010)


6

2D point sequences

a b c d e f g h i j

Build double linked list



Toolkit, 2010)


6

2D point sequences

ab c d ef g h ij

AHC



Toolkit, 2010)


6

2D point sequences

ab c d ef g h ij

AHC



Toolkit, 2010)


6

2D point sequences

ab efg ij

Extract line segments


• Non-trivial Generalization to 3D

7



– None-overlapping nodes

7




7




– Number of merging attempts

7

ef g h

≤2




– Number of merging attempts

7

ef g h ab efg h

c

d ijk

≤2 ?


8

• Algorithm Overview

Agglomerative Hierarchical Clustering

(a) Initialize graph

Repeat if merging MSE ≤ threshold

(c) Merge with neighbor node B which gives min merging MSE

B

A(d) Extract Coarse Planes

Otherwise don’t merge but extract A

(e) Refine details(b) Find node Awith min MSE

A


8

• Algorithm Overview


• Graph Initialization

9



– Non-overlapping node initialization

9




– Rejecting “bad” nodes

9





9

1) High MSE





9

2) Missing Data





9

3) Depth Discontinuities





9

4) At Boundary Between Planes





– Good! Avoid per-point normal estimation

9


• Agglomerative Hierarchical Clustering

10



10



10

Current min MSE node

1 Cluster Step(s)



10

Current min MSE node

Best node to merge

1 Cluster Step(s)



10

2 Cluster Step(s)



10

3 Cluster Step(s)



10

300 Cluster Step(s)



10

472 Cluster Step(s)


• Average Number of Merging Attempts

11

ab efg h

c

d ijk


– Small irrespective of initial number of nodes

11

ab efg h

c

d ijk


– Planar graph! Average node degree < 6

11

ab efg h

c

d ijk


– Merging is empirically a constant-time operation

• O(nlogn), only arise from maintaining the min-heap

11

ab efg h

c

d ijk


• Implementation Details

12



– Disjoint set

12



– Disjoint set

– Min-heap

12



– Disjoint set

– Min-heap

– Second-order statistics

12



– Disjoint set

– Min-heap


– Depth discontinuity/MSE threshold (Holzer et al. IROS 2012; Khoshelham & Elberink, 2012)

12



– Disjoint set

– Min-heap


– Depth discontinuity/MSE threshold (Holzer et al. IROS 2012; Khoshelham & Elberink, 2012)

– Avoid strip-like initial node shape

12


• Segmentation Refinement

13



– Artifacts

13

Sawtooth



– Artifacts

13

Unused Data



– Artifacts

13

Over-segmentation



– Artifacts

– Pixel-wise region-grow refinement

• Only check boundary blocks and points

13


– Pixel-wise region-grow refinement

• Only check boundary blocks and points

13


• Simulated Data

– Robustness to uniformly distributed depth noise (Georgiev et al., IROS 2011)

– Noise magnitude E = 0, 10, …, 200mm

– Ground truth depth ranges from 1396mm to 3704mm

14

E=150E=0 E=200E=100


• Real-World Kinect Data

– 2102 frames of an indoor scene

– 640 × 480 pixel/frame

15

Initial node size 10x10





– 2102 frames of an indoor scene

– 640 × 480 pixel/frame

15

27.3 ± 6.9ms/frame > 35Hz



16


• SegComp Datasets (Hoover et al. PAMI 1996)

– ABW-TEST

– PERCEPTRON-TEST

17


• SegComp Benchmark (Gotardo et al. CVPR 2003; Oehler et al.

ICIRA 2011; Holz & Behnke IAS 2012)

18


• We presented an efficient plane extraction algorithm based on agglomerative clustering for organized point clouds.

19



• We analyzed the complexity of the clustering algorithm and shown that it is log-linear in the number of initial nodes.

19



• We analyzed the complexity of the clustering algorithm and shown that it is log-linear in the number of initial nodes.

• We demonstrated real-time performance with the accuracy comparable to state-of-the-art algorithms.

19


• Acknowledgement

– MERL

• Jay Thornton

• Srikumar Ramalingam

– Rackham Graduate School, University of Michigan

– NSF Student Travel Grant

20

21

Chen Feng (in Chinese: 冯晨) PhD Candidate Department of Civil and Environmental Engineering University of Michigan, Ann Arbor E-mail: [email protected] Web: http://www.umich.edu/~cforrest/

mailto:[email protected]

http://www-personal.umich.edu/~cforrest/



Documents

Fast Plane Extraction in Organized Point Clouds Using … · 2019. 10. 22. · Fast Plane Extraction in Organized Point Clouds Using Agglomerative Hierarchical Clustering Chen Feng1),