This thesis describes the research that was performed during the design and construction of an active vision system capable of autonomous 3D scene reconstruction. Stereo-pairs of images are captured from calibrated steerable cameras mounted upon a previously constructed binocular robot head, the ASP. Range data from which three-dimensional points are derived is constructed through photogrammetric analysis of the images captured. The correspondence problem that forms part of the photogrammetric analysis is solved by stereo-matching algorithms. Frame-rate performance by these algorithms has only previously been achieved on standard computational hardware through trade-offs against quality. It is argued that foveation, a method of data-reduction found in mammals is an appropriate mechanism to improve performance without noticeable loss of quality. Foveation is a form of multi-resolution data analysis, where processing effort is concentrated on points of interest. A foveated stereo-matching algorithm is derived from a contemporary algorithm, implemented and characterized. The concept of foveation and accompanying multi-resolution analysis as a legitimate method to improve performance is demonstrated as further components of the active vision system are created. Potential fixation points, points to be imaged at the highest possible spatial resolution are selected by multi-resolution analysis of the acquired range data. A three-step approach based upon sampling resolution is applied to surface registration, required to correctly align the individual 3D models. This greatly reduces the computational effort required to align models from minutes to under a second. A vergence algorithm was required to ensure that the fields of view of both cameras overlap, a requirement for 3D data to be extracted. A novel three-step approach is proposed and implemented, based upon the foveated stereo-matching algorithm derived. The approach is designed to position the cameras to minimize the high-resolution error recovered when matching image pairs with the foveated matching algorithm. Finally, the individual components are successfully integrated into a unified system. The validity of the approach is demonstrated as a multi-resolution 3D model of a test subject is recovered.