Robot Vision

Group head: Professor John Illingworth

The emphasis of our research is on active `looking' rather than passive `seeing'. Visual understanding is an evolving description gathered over time. We wish to explore the question of what can be achieved over time by control of all aspects of the vision process i.e. how can adjustment of image sensors produce better data, how can choice and control of image analysis algorithm yield improved information and finally what is the most appropriate system response given the desired task and prior (but incomplete and uncertain) knowledge of the environment. To address this ambitious research programme, work is in progress on several related fronts including fundamental research in statistical estimation, geometry of three-dimensional structures and structure-from-motion. New theories and algorithms are validated on a range of mechanical platforms including mobile robot platforms and a precision robot arm.

Mobile Robot Localisation

Robots need a sense of their location if spatial information is to be integrated over time. This can be done using odometry information i.e. estimates of motion made using on-board wheel or gyroscopic sensors. However, this is often error prone due to effects such as wheel slippage or tyre wear. This project explores how machine vision can supply an independent estimate of robot motion to confirm odometry as well as exploring well founded ways to combine vision and odometry estimates, taking account of their differing characteristics.

Stable Representation of Features in Images

The localisation and tracking of image features is fundamental for reconstruction. Features can also be used to guide reactive behaviour for applications such as collision avoidance or docking. Our localisation approach exploits geometric constraints in the 3D environment to infer the position of features in consecutive frames. Features define relative motion and are then used to program the robot to move in a way that it obtains a predetermined image.

Robust Map Building and Motion Planning

The aim of this project is to develop a robot that can explore an unknown space and incrementally construct a model of that environment. For simple tasks such as moving whilst avoiding obstacles, approaches based around simple features and occupancy maps have proven effective. However, for fuller descriptions correspondingly more sophisticated methods of extracting full 3D surface information have been developed. These methods exploit state of the art techniques such as Variable State Dimension Filters.