Vision-based localisation is a biologically-inspired process.  In contrast to the development of robotic research in the past few decades, only a few types of nocturnal animals, such as bats, use direct range sensing as a mean to explore the environment.  The heavy computation requirement involved in near real-time image processing hindered the widespread use of vision in robotics.  Due to the continuous advances in electronics, this physical barrier has largely been overcome nowadays.  It is now about time for us to explore further into this fascinating area.

To be more specific, statistical vision-based localisation is my research focus.  If we want the robot to localise using natural landmarks extracted from the scene, it will have to match these landmarks before carrying out triangulation.  Image features are not unique in many scenes, which may be leaves on a tree or identical chairs in a conference room.   It is different to match these similar objects precisely.  Robust estimation methods, e.g. RANSAC, can be improved to take advantage of the extra available information.  Panoramic image geometry has a number of desirable properties, including very large field-of-view and ease processing, which is particularly relevant for robotics.  Its use has been integrated to my robot localisation algorithm.

Recent development in Sequential Monte Carlo (SMC) methods offers very efficient estimation to the Bayes' Theorem.   Many different classes of localisation problems, including absolute position estimation, tracking and the very difficult kidnapped robot problem, have been solved using SMC.  Possible extension to localisation problems, in which neither robot nor landmark positions are known, is one of the other direction that I am researching on at this moment.  Research to extend the algorithm for localisation using image features is still on-going.

For more details:

• Selected presentations
• Publications