Research objectives and content
One of the principal aims of computer vision is to develop automated techniques for understanding 3D properties of the environment given video imagery obtained from one or more cameras. Of considerable recent interest has been to what extent information can be inferred about the world when the camera parameters, camera orientations and camera motions are unknown. In this situation, we need to embark upon a preprocess that attempts to recover these factors involved in the picture taking process, prior to computing 3D descriptions of the scene. The process of automated determination of some of the intrinsic and extrinsic camera parameters is termed self-calibration; it forms the basis of the proposed project. In this project, the issue of self-calibration in the special context of a stereo head, perhaps the most commonly adopted binocular camera configuration in robotics, will be revisited. In previous work, closed form solutions were obtained for the calibration parameters and for the ego-motion parameters of the stereo platform, when it undergoes ground plane motion. Key factors of the methodology used for this project will be the consideration of explicit, analytical forms of the fundamental matrix, and the use of the stereo-head state feedback information, provided by the electro-mechanical active stereo platform.
Specifically, the objectives of this research project will be the following. To utilise the proposed methodology, based on the study of the analytical form of the fundamental matrix, to extend the self-calibration of the stereo platform to more genera forms of motion. The
self-calibration problem when the intrinsic parameters are not fixed between views but free to vary will also be studied. This will permit the cameras to change their internal parameters such as their focal lengths. When the information concerning the relative orientation of the stereo cameras is known, the intrinsic parameters of the cameras may be obtained solely from point correspondences, or directly from the fundamental matrix coefficients.
Training content (objective, benefit and expected impact)
The relevance of this research project is apparent in that it contributes to an understanding on how visual information may be used in order to infer 3D properties of the world. More specifically this project aims to make an important contribution to the problem of dealing with uncalibrated cameras, a relatively new endeavour in computer vision. In particular, the design of new self calibration techniques in the special context of an active stereo platform the most popular configuration in robotics, will be pursued. The training of the applicant will be most benefited from the study of the problem of recovering 3D structure when both the motion of the cameras and their internal parameters are unknown.
Links with industry / industrial relevance (22)
The applications of uncalibrated vision to industrial applications are numerous. The advances in this field may be applied to surveillance applications, to assist in teleoperation, to produce stereo visualisations of objects which may assist in non invasive surgery, interior design, autonomous vehicle navigation and many others. The Robotics Research Group has extensive
collaborations with industries, involving 20 companies from different countries.