Design of Stereo Heads

The introduction of definite and simple action goals into machine vision at an early stage promises to address the defects which have become apparent in the data-driven reconstructionist paradigm, with its relentless emphasis on accurate geometric recovery of the whole 3D scene — a reconstruction which is often unnecessary and usually too difficult. Simple goals allow a more precise definition of what is salient in the scene for the particular task in hand, in turn indicating what image data are relevant, and allowing the sensors to be directed to view the important parts of the scene. One benefit of having control over where and how the scene is observed over time is that tasks which are ill-posed in the static case may become well-posed [1]. A second is that finite computational resources can be allocated where they are needed, and yet a third is that passive vision becomes an active pointing device. Such purposive and active vision promises to open up exciting new avenues of experiment, in which the researcher has to relinquish precise control over scene conditions and the behaviour of the vision system. But to reach this experimental Elysium, the vision researcher needs apparatus; not only an accurate mechanical device to steer the cameras, but also a high bandwidth closed-loop vision sensor and controller. He immediately encounters other significant problems more familiar to control engineers: What is an appropriate mechanism? How can the time/speed requirements of the task be met? What is the control state, and observer and control law? And so on. In this paper we outline the design considerations behind our construction of a steerable stereo head/eye platform, the mechanical design and that of the gaze controller and vision systems. We show how the kinematics of the platform can be used to relate image measurements with the scene motion in a static frame.