Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Large-scale 3D Mapping of Subarctic Forests
Philippe Babin, Philippe Dandurand, Vladimír Kubelka, Philippe Giguère and François Pomerleau
12th FSR conference, Tokyo, 2019
Subarctic Boreal Forest: Research Opportunity
2/27
Our approachNaive approach
3/27
Applications
4/27
Local WildlifeSnow Fall Path obstacles Uneven Path
Challenge of Field Tests
5/27
Mapping of Subarctic Boreal Forest -
Challenges
Unstructured environment → hard to map
Cold temperatures → noisy sensor
Few visual features due to snow → bad for vision based approaches
6/27
Related Work
Williams et al., 2009 Paton et al., 2016
7/27
Contributions
Large-scale mapping of difficult environments
Novel fusion of IMU and GNSS measurement inside of ICP
Generated maps are crisp and without long term drifts
Introduced optimization to scale to large map
8/27
Dataset Environment
4.1 km of forest path
9/27
10/27
Data Acquisition Platform
GNSS station (RTK)
RS-16 lidar
MTI-30 IMU
10h of battery life
11/27
Iterative Closest Point (ICP)
T ?
12/27
ICPTinit
T
13/27
Iterative Closest Point (ICP)
Iterative Closest Point (ICP)
14/27
Sensor Fusion
Lidar ICP
IMU
GNSS
SLAM
Lidar
Penalty-ICPIMU
GNSS
mappose
Classical approach Our approach
mappose
Covariance [1]pose
[1] D. Landry, F. Pomerleau, and P. Giguère. CELLO-3D: Estimating the Covariance of ICP in the Real World. In ICRA, 2019
15/27
Point cloud
Pose
Covariance
Pose/Covariance
Legend
Prior ICP with penalty ICP no penalty
Map of lake Dataset
350m
16/27
ICP with penalty ICP no penalty
17/27
Crispiness
locally consistent
Prior ICP With penalties ICP Without Penalties
18/27
Prior ICP with penalty ICP no penalty
Map of forest Dataset
500m
19/27
Prior ICP with penalty ICP no penalty
Map of forest Dataset
500m
20/27
Map of skidoo Dataset
ICP no penaltyICP with penaltyPrior
670m
21/27
ICP no penaltyICP with penaltyPrior
670m
22/27
Map of skidoo Dataset
Full Map with Penalty
23/27
Performance improvements
24/27
Future work
25/27
Future work – Project SNOW
26/27
Questions?
27/27
28
Performance improvements
29/26
Results: effect of penalties
30
Penalty-ICP• Leverage ICP’s minimizer for sensor fusion
• Add penalty term based on GNSS and IMU estimate
• Introduced a point to Gaussian cost function
• Minimize Mahalanobis distance instead of the Euclidian distance
31
Full Map
32