Robot SLAM and navigation with multi-camera computer vision

In this thesis we focus on computer vision capabilities suitable for practical massmarket mobile robots, with an emphasis on techniques using rigs of multiple standard cameras rather than more specialised sensors. We analyse the state of the art of service robotics, and attempt to distill the vision capabilities which will be required of mobile robots over the mid and long-term future to permit autonomous localisation, mapping and navigation while integrating with other task-based vision requirements. The first main novel contribution of the work is to consider how an ad-hoc multicamera rig can be used as the basis for metric navigation competences such as featurebased Simultaneous Localisation and Mapping (SLAM). The key requirement for the use of such techniques with multiple cameras is accurate calibration of the locations of the cameras as mounted on the robot. This is a challenging problem, since we consider the general case where the cameras might be mounted all around the robot with arbitrary 3D locations and orientations, and may have fields of view which do not intersect. In the second main part of the thesis, we move away from the idea that all cameras should contribute in a uniform manner to a single consistent metric representation, inspired by recent work on SLAM systems which have demonstrated impressive performance by a combination of off-the-shelf or simple techniques which we generally categorise by the term ‘lightweight’. We develop a multi-camera mobile robot vision system which goes beyond pure localisation and SLAM to permit fully autonomous mapping navigation within a cluttered room, requiring free-space mapping and obstacle-avoiding planning capabilities. In the last part of the work we investigate the trade-offs involved in defining a camera rig suitable for this type of vision system and perform some experiments on camera placement.