Structure from Motion and Pursuit-Evasion Games

Rene Vidal


Abstract

Computer vision has been studied over the past 20 years and many algorithms have been proposed in structure from motion, stereo, segmentation, shape, illumination, etc. But how many of these algorithms are being used for real-time control of cars, helicopters, planes, robots, etc.? Unfortunatelly not too many.

This is what motivated us to study computer vision, and structure from motion (SFM) in particular, from a control-theoretic perspective. We have done so within the Berkeley Aerial Robot Project, in which a number of ground and aerial robots form a team of pursuers that tries to capture a team of evaders in an unknown environment. The main vision tasks we want to acomplish are to detect position, orientation and velocities of the evaders; to detect position and orientation of obstacles; to estimate position and orientation of a moving helicopter; and to use all this information for navigation and control purposes such as pursuit evasion games, vision based landing of a helicopter, formation flight, etc.

In the talk, I will first describe our recent results in Single-Body SFM:
1) Multiple View Geometry
2) Implementation of Pursuit-Evasion Games.

I will then talk about our future research in Multi-Body SFM, the challenges we will have to face, and the way we plan to address the problem.


Maintained by Philippos Mordohai