4. CROP MONITORING INSPECTION of INFRASTRUCTURE
DIGITIZATION IN ARCHAEOLOGY SEARCH & RESCUE
Computer
Vision
&
Robotics
5. Unpacking the Idea(see notes for details)
[IROS 2012] with M. Achtelik, S. Lynen, S. Weiss, L. Kneip and R. Siegwart
6. Teaching Robots to See – are we there yet?
Handle larger amounts of data more effectively
§ Competing goals:
Precision vs. Efficiency
• Fast motion
• Large scales
• Rich maps
• Robustness
• Lower Computation
1ms Target Tracking Vision Chip by Ishikawa
Komuro University
Photo credits: Max Pixel
7. Weight
ü Lighter & safer than larger robots
✗ Limited resources: sensors, computation, …
Agility
ü Very agile, fast
✗ Fast, unstable dynamics, cannot “stop”
Autonomy
✗ Battery, communication bandwidth
Teaching Drones to See – properties & challenges
Photo credit: Francois Pomerleau
11. … e.g. via sensor fusion
[ICRA 2018]: with R. Mascaro, L. Teixeira, T.
Hinzmann and R. Siegwart
?
1
12. Localization & Mapping downtown Zurich
Estimated Trajectory
Estimated Map
Current Map points
Place Recognition
• Recognize when the robot visits a
place it has seen before.
• Search for matching images across
robot’s trajectory.
Camera view
?
1
13. Place Recognition
Different places can appear identical
Large viewpoint changes (especially from an aircraft)
Seasonal / Illumination changes
Place appearance changes between visits
14. [3DV 2017] with F. Maffra,
L. Teixeira and Z. Chen
Place Recognition using orthophotos
Current view
Current Orthophoto
Database match: Original view
Database match: Orthophoto
?
1
15. 1
?
Current View Database match
[IROS 2017]: with Z. Chen, F. Maffra and I. Sa
Place Recognition using deep learning