Presentation is loading. Please wait.

Presentation is loading. Please wait.

May 16, 2015 Sparse Surface Adjustment M. Ruhnke, R. Kümmerle, G. Grisetti, W. Burgard.

Similar presentations


Presentation on theme: "May 16, 2015 Sparse Surface Adjustment M. Ruhnke, R. Kümmerle, G. Grisetti, W. Burgard."— Presentation transcript:

1 May 16, 2015 Sparse Surface Adjustment M. Ruhnke, R. Kümmerle, G. Grisetti, W. Burgard

2 Metric 3D Models ► Essential for tasks like: ► Object recognition ► Manipulation

3 Metric 3D Models ► Essential for tasks like: ► Object recognition ► Manipulation ► Key challenges in model acquisition with mobile robots ► Errors in pose estimate ► Measurement errors

4 Model Creation ► Optimize the sensor poses ► Registration / SLAM ► Reduce the impact of measurement errors ► Use optimal sensor distance ► Local noise reduction techniques (Moving Least Squares, Statistical Outlier Removal, …) ► Pose information is mostly not considered ► Sensor pose gives information about normal direction and range of the measurement

5 Sparse Surface Adjustment ► Goal: Jointly optimize robot poses and surface points positions ► Surface Model ► Model of measurement uncertainties ► Data association: find corresponding points ► Utilize sparse graph optimizer framework g 2 o

6 Surface Model ► Range measurements sample surfaces ► Assumption: Piecewise regular surfaces ► Surface sample ► 3D Position ► covariance ► normal (local neighborhood)

7 ► Range measurement ► Sensor specific Covariance ► Dependent on range and incidence angle ► Gaussian error distributions Sensor Model front view side view ~ 3.5m ~ 0.7m Kinect RGB-D

8 Data Association ► Normal shooting as data association heuristic ► Assign surfaces samples of different observations ► Covariance ► Large error weight in direction of the normal ► Small weight for errors in tangential direction

9 Optimization ► Iteratively: ► Optimize system with g 2 o ► Re-compute: ► surface point characteristics (covariance, normals) ► data association

10 SSA 2D: Intel Dataset

11 Object 5mm Resolution

12 AASS Loop Dataset* SLAM result (input) SSA result *Courtesy of Martin Magnusson, AASS, Örebro, Sweden

13 SSA 3D: Example Cup

14 Example: Scan Refinement ► SSA refines scans based on more certain nearby measurements after optimizationraw scan

15 Comparison SSA / MLS ► Moving Least Squares (MLS) ► Local smoothing method ► No correction of robot poses ► Sparse Surface Adjustment (SSA) ► Robot pose correction & smoother surfaces SSA result MLS result

16 Summary SSA ► Iterative refinement of ► Sensor poses ► Surface points positions ► Considering range & sensor dependent uncertainties ► Re-computation of data association ► Uses PCL and FLANN


Download ppt "May 16, 2015 Sparse Surface Adjustment M. Ruhnke, R. Kümmerle, G. Grisetti, W. Burgard."

Similar presentations


Ads by Google