Towards Fully Autonomous Visual Navigation

This thesis addresses some key issues which affect the level of autonomy inherent in visual navigation systems, with wider applicability in a range of fields. They can be divided into two areas. Firstly, automated initialisation, in which the kinematic and camera calibration parameters needed for an active camera platform are calculated without user interaction. Secondly, the complexity problem of simultaneous localisation and mapping using the Extended Kalman Filter, which is a highly general and flexible localisation methodology for keeping track of the position of a vehicle as it navigates an unknown scene. The problem, which is inherent in any probabilistic mapping filter, is that the computational cost of incorporating new measurements scales with the size of the region being explored. Alignment is required for any active camera used to make measurement by way of angle encoders. This is the process of bringing each axis to a natural origin defined by an active head’s kinematics, and is necessary for referring measurements back to a fixed coordinate frame. In this thesis, alignment is achieved by detecting image-based invariants to motions about individual axes. A number of algorithms are developed for most typical scenarios, relying only on scene rigidity and the ability to control head axes individually. Camera calibration allows metric information about the scene to be deduced from the camera images. This thesis develops a number of new tools for self-calibration of an active camera, again reliant on scene rigidity and head control. In particular, some closed-form solutions to the difficult problem of calibrating in a planar scene are provided for the binocular and monocular cases, which do not involve calculations on reconstructed scene data. Finally, the issue of complexity is addressed in two ways. Firstly, the requirements for a successfully operating localisation filter are established, and the ability of the Extended Kalman Filter to achieve these investigated. Based on this analysis, a proposal is made for a solution to the problem, involving making optimal transitions between map partitions. Secondly, the process of Postponement is derived and developed. Postponement, and its logical extension Hierarchical Postponement, allow the maximum efficiency to be extracted from the EKF by restricting computation to the update of those features being measured. This is done without sacrificing the ability to bring any part of the map fully and consistently up to date if required. Using Dynamic Submapping the set of features being maintained can be defined on the fly, and allows costly updates of the global map to be delayed until there is processor time available to carry them out, while the majority of filter steps have a bounded cost. Using Hierarchical Postponement, updates can be permanently restricted to a small top level map. All these tools are applicable to fields outside visual navigation: alignment to measurement using active cameras, calibration to computer vision in general, and efficient EKF SLAM to any such system regardless of the measurement methodology.