Probabilistic Surfel Fusion for Dense LiDAR Mapping Chanoh Park, Soohwan Kim, Peyman Moghadam, Clinton Fookes, Sridha Sridharan

Transcription

1 Probabilistic Surfel Fusion for Dense LiDAR Mapping Chanoh Park, Soohwan Kim, Peyman Moghadam, Clinton Fookes, Sridha Sridharan ICCV workshop

2 Introduction Narrow FoV Short sensing range Sensitive to ambient light This work covers the map fusion for LiDAR 2

3 Introduction What is a surfel? Surface + Element Points Cloud map Surfels map 3

4 Introduction What make it difficult to introduce surfel fusion in LiDAR-based SLAM Absence of projective data association Radius search Uncertainty based search 4

5 Introduction What make it difficult to introduce surfel fusion in LiDAR-based SLAM Absence of projective data association Existence of the surfel degeneracy Radius search Uncertainty based search Unstable normal direction 5

6 Overview Local Mapping Global Mapping Multi-Resolution Sparse Surfels Localization and Surfel Integration Active Area Map Update Sparse Surfel Map Raw Points Cloud Local SLAM Module Transformation LiDAR Dense Surfels Dense Surfel Fusion Radius Search Map Update Dense Surfel Map 6

7 Dual Map Representation Robust ICP Fast Visualization Too sparse 3D Ellipsoidal Surfel Map from Multi-resolutional Voxel Hassing 2D Disk Surfel Map from Nearest Neighbor Searching *Color is coded with normal direction 7

8 Dual Map Representation Local Mapping Global Mapping Multi-Resolution Sparse Surfels Localization and Surfel Integration Active Area Map Update Sparse Surfel Map Raw Points Cloud Local SLAM Module Transformation by Point-to-Plane ICP LiDAR Dense Surfels Dense Surfel Fusion Radius Search Map Update Dense Surfel Map 8

9 Noise Modelling Distance Surfel LiDAR Beam Incident angle Surfel normal 3D position uncertainty Surfel position uncertainty modeling 9

10 Noise Modelling Possible distribution of normal vector Surfel Beam LiDAR Surfel normal Surfel normal uncertainty Surfel normal direction uncertainty modeling 10

11 Dense Surfel Matching Local Mapping Global Mapping Multi-Resolution Sparse Surfels Localization and Surfel Integration Active Area Map Update Sparse Surfel Map Raw Points Cloud Local SLAM Module Transformation LiDAR Dense Surfels Dense Surfel Fusion Radius Search Map Update Dense Surfel Map Merge 11

12 Dense Surfel Matching Global Map Surfels Local map Matching candidates Global map 12

13 Dense Surfel Matching Global Map Surfels Step1: Resolutional Matching r3 r1 r2 r4 Local map real structure Side view of the matching candidates Matching candidates Global map 13

14 Dense Surfel Matching Global Map Surfels Step1: Resolutional Matching r3 r1 r2 r4 Local map real structure Side view of the matching candidates Matching candidates Step2: Depth Matching Uncertainty along the normal direction σ 2 Global map real structure 14

15 Dense Surfel Matching Benefits of this approach 1. Easy control of the surface resolution Space digitization is not required 2. It searches more in the laser beam direction Better noise handling Higher Resolution with a Euclidian distance threshold Deeper searching with a Mahalanobis distance threshold 15

16 Dense Surfel Fusion Black line : fused normal uncertainty on the unit sphere Fusion Fusion Unit sphere surface Tangentiality reinforcement Uncertainty Normal Before After 16

17 Experiment Results 17

18 Real Data Experiment Trajectory Utilized scanning system Surfel map with normal direction color coded Color fused surfel map 18

19 Real Data Experiment CT-SLAM Proposed 0.7m patch extracted from the maps *Position error is in mm. Normal error is in degree. 19

20 Real Data Experiment 20

21 Summary 1. Probabilistic dense surfel fusion for LiDAR is proposed 2. Our method shows denser but lesser noise level 3. An advantage on long-term SLAM applications 21

22 Questions?