1. Asian Institute of Technology
Indoor SLAM Monocular SLAM with indoor surface estimation and visualization
Where are you?
Umm not sure..
Thesis by:
Ran Zask
Asian Institute of Technology

2. Agenda Background Problem statement Related work Our algorithm Results
3D modeling
Texture mapping
Results
Problems & limitations
Conclusions & Recommendations

3. Background
Some robot applications require human operator (e.g. search & rescue)
Human operator needs to maintain situation awareness in order to control the robot.
Current means of maintaining situation awareness are poor (2D maps, live video)
3D modeling and visualization is needed.

4. Problem statement
Real-time SLAM implementations do not estimate surfaces of the environment and using expensive sensors (sonar, laser)
Structure From Motion is too heavy and done offline.
Real-time 3D modeling and visualization is feasible and needed.

5. Related work Pollefeys et al. (2004) : point-based modeling
Monocular camera
3D reconstruction from feature points
Bundle adjustment
Rectification, Dense stereo
Offline algorithm, Massive computation
Similar application: Microsoft Photosynth

6. Related work (cont.) Iterative closest point (ICP) Mesh registration
Johnson and Kang (1999) : grid-based modeling
Omni-directional stereo
2D Delauney triangulation
Iterative closest point (ICP)
Mesh registration
Massive computation

7. The algorithm Single camera, calibrated Requires SLAM (point-based)
Uses 3D points and occupancy grids
Online algorithm
Creates 3D models with texture mapping
Provides low metric accuracy
Less noise between grid cells

8. The algorithm Uses 2 occupancy grids: Local grid Global grid
Recreated each frame
Contains ‘fullness’ seen at the current frame
Incrementally added to the global grid.
Global grid
Contains the incremental ‘fullness’
Used for Iso-surface calculation

9. The algorithm (3D modeling)
For each image:
Input: new 3D point set (from SLAM)
Reset local grid
Associate 3D points with cells in the local grid
Ignore cells associated by few points
Add camera-center as an occupied cell
Fill-in the grid (convex-hull)
Merge local grid into global grid
Iso-surface the global grid -> last model

11. The algorithm (texture mapping)
For each image :
(After having the latest model)
Classify triangles as: “new,” “old,” and “expired”
Project “new” triangles onto current frame -> textmaps
Remove “expired” triangles

12. Results Modeling accuracy is satisfying Texture mapping
Tessellation within a patch: perfect
Tessellation between patches: Inaccurate
The more images are closer to each other, the better the modeling and the texture mapping
“patch” – collection of several triangles of the same image

13. Modeling results

14. Modeling results

15. Modeling results

16. Summary
We created an online algorithm for modeling with texture, all from a single camera.
Low computation is required per image
This should allow short baseline -> good results
Real time implementation of the system is feasible.
Why it’s a useful aproach

17. Future work Create a real-time system
Create 3D tools for the human operator to easily ‘walk’ in the model and navigate the robot.
Improving the algorithm even more:
performing iso-surface on tighter grids
Why it’s a useful aproach

18. DEMO